From 11dbac57e92e8ecfc1f064193228f1ad9afaf303 Mon Sep 17 00:00:00 2001
From: SpeedProg <speedprog@googlemail.com>
Date: Thu, 02 Jan 2020 16:04:02 +0000
Subject: [PATCH] removed weights
---
opencv_dnn.py | 852 ++++++++++++++++++++++++++++++++++----------------------
1 files changed, 514 insertions(+), 338 deletions(-)
diff --git a/opencv_dnn.py b/opencv_dnn.py
old mode 100644
new mode 100755
index 03d4cc5..4b2c4e0
--- a/opencv_dnn.py
+++ b/opencv_dnn.py
@@ -1,34 +1,43 @@
-import cv2
-import numpy as np
-import pandas as pd
-import imagehash as ih
-import os
+import argparse
import ast
-import sys
-import math
-import random
import collections
+import cv2
+import imagehash as ih
+import numpy as np
from operator import itemgetter
-import time
+import os
+import pandas as pd
from PIL import Image
+import time
+from multiprocessing import Pool
+from config import Config
import fetch_data
-import transform_data
-
-card_width = 315
-card_height = 440
-def calc_image_hashes(card_pool, save_to=None, hash_size=32, highfreq_factor=4):
+"""
+As of the current version, the YOLO network has been removed from this code during optimization.
+It was found out that YOLO was adding too much processing delay, and the benefits from using it couldn't justify
+such heavy cost.
+If you're interested to see the implementation using YOLO, please check out the previous commit:
+https://github.com/hj3yoo/mtg_card_detector/tree/dea64611730c84a59c711c61f7f80948f82bcd31
+"""
+
+def do_calc(args):
+ card_pool = args[0]
+ hash_size = args[1]
new_pool = pd.DataFrame(columns=list(card_pool.columns.values))
- new_pool['card_hash'] = np.NaN
- new_pool['art_hash'] = np.NaN
+ for hs in hash_size:
+ new_pool['card_hash_%d' % hs] = np.NaN
+ new_pool['set_hash_%d' % 64] = np.NaN
+ #new_pool['art_hash_%d' % hs] = np.NaN
for ind, card_info in card_pool.iterrows():
if ind % 100 == 0:
- print(ind)
+ print('Calculating hashes: %dth card' % ind)
card_names = []
+ # Double-faced cards have a different json format than normal cards
if card_info['layout'] in ['transform', 'double_faced_token']:
- if isinstance(card_info['card_faces'], str): # For some reason, dict isn't being parsed in the previous step
+ if isinstance(card_info['card_faces'], str):
card_faces = ast.literal_eval(card_info['card_faces'])
else:
card_faces = card_info['card_faces']
@@ -38,30 +47,84 @@
card_names.append(card_info['name'])
for card_name in card_names:
+ # Fetch the image - name can be found based on the card's information
card_info['name'] = card_name
- img_name = '%s/card_img/png/%s/%s_%s.png' % (transform_data.data_dir, card_info['set'],
+ cname = card_name
+ if cname == 'con':
+ cname == 'con__'
+ img_name = '%s/card_img/png/%s/%s_%s.png' % (Config.data_dir, card_info['set'],
card_info['collector_number'],
- fetch_data.get_valid_filename(card_info['name']))
+ fetch_data.get_valid_filename(cname))
card_img = cv2.imread(img_name)
+
+ # If the image doesn't exist, download it from the URL
if card_img is None:
+ set_name = card_info['set']
+ if set_name == 'con':
+ set_name = 'con__'
fetch_data.fetch_card_image(card_info,
- out_dir='%s/card_img/png/%s' % (transform_data.data_dir, card_info['set']))
+ out_dir='%s/card_img/png/%s' % (Config.data_dir, set_name))
card_img = cv2.imread(img_name)
if card_img is None:
print('WARNING: card %s is not found!' % img_name)
+ continue
+ """
+ img_cc = cv2.cvtColor(card_img, cv2.COLOR_BGR2GRAY)
+ img_thresh = cv2.adaptiveThreshold(img_cc, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY_INV, 11, 5)
+ # Dilute the image, then erode them to remove minor noises
+ kernel = np.ones((3, 3), np.uint8)
+ img_dilate = cv2.dilate(img_thresh, kernel, iterations=1)
+ img_erode = cv2.erode(img_dilate, kernel, iterations=1)
+ cnts, hier = cv2.findContours(img_erode, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
+ cnts2 = sorted(cnts, key=cv2.contourArea, reverse=True)
+ cnts2 = cnts2[:10]
+ if True:
+ cv2.drawContours(img_cc, cnts2, -1, (0, 255, 0), 3)
+ #cv2.imshow('Contours', card_img)
+ #cv2.waitKey(10000)
+ """
+ set_img = card_img[595:635, 600:690]
+ #cv2.imshow(card_info['name'], set_img)
+ # Compute value of the card's perceptual hash, then store it to the database
#img_art = Image.fromarray(card_img[121:580, 63:685]) # For 745*1040 size card image
- #art_hash = ih.phash(img_art, hash_size=32, highfreq_factor=4)
- #card_pool.at[ind, 'art_hash'] = art_hash
img_card = Image.fromarray(card_img)
- card_hash = ih.phash(img_card, hash_size=hash_size, highfreq_factor=highfreq_factor)
- #card_pool.at[ind, 'card_hash'] = card_hash
- card_info['card_hash'] = card_hash
- #print(new_pool.index.max())
+ img_set = Image.fromarray(set_img)
+ #cv2.imshow('Set' + card_names[0], set_img)
+ for hs in hash_size:
+ card_hash = ih.phash(img_card, hash_size=hs)
+ set_hash = ih.whash(img_set, hash_size=64)
+ card_info['card_hash_%d' % hs] = card_hash
+ card_info['set_hash_%d' % 64] = set_hash
+ #print('Setting set_hash_%d' % hs)
+ #art_hash = ih.phash(img_art, hash_size=hs)
+ #card_info['art_hash_%d' % hs] = art_hash
new_pool.loc[0 if new_pool.empty else new_pool.index.max() + 1] = card_info
+ return new_pool
- new_pool = new_pool[['artist', 'border_color', 'collector_number', 'color_identity', 'colors', 'flavor_text',
- 'image_uris', 'mana_cost', 'legalities', 'name', 'oracle_text', 'rarity', 'type_line',
- 'set', 'set_name', 'power', 'toughness', 'art_hash', 'card_hash']]
+def calc_image_hashes(card_pool, save_to=None, hash_size=None):
+ """
+ Calculate perceptual hash (pHash) value for each cards in the database, then store them if needed
+ :param card_pool: pandas dataframe containing all card information
+ :param save_to: path for the pickle file to be saved
+ :param hash_size: param for pHash algorithm
+ :return: pandas dataframe
+ """
+ if hash_size is None:
+ hash_size = [16, 32]
+ elif isinstance(hash_size, int):
+ hash_size = [hash_size]
+
+ num_cores = 16
+ num_partitions = round(card_pool.shape[0]/1000)
+ if num_partitions < min(num_cores, card_pool.shape[0]):
+ num_partitions = min(num_cores, card_pool.shape[0])
+ pool = Pool(num_cores)
+ df_split = np.array_split(card_pool, num_partitions)
+ new_pool = pd.concat(pool.map(do_calc, [(split, hash_size) for split in df_split]))
+ pool.close()
+ pool.join()
+ # Since some double-faced cards may result in two different cards, create a new dataframe to store the result
+
if save_to is not None:
new_pool.to_pickle(save_to)
return new_pool
@@ -69,10 +132,12 @@
# www.pyimagesearch.com/2014/08/25/4-point-opencv-getperspective-transform-example/
def order_points(pts):
- # initialzie a list of coordinates that will be ordered
- # such that the first entry in the list is the top-left,
- # the second entry is the top-right, the third is the
- # bottom-right, and the fourth is the bottom-left
+ """
+ initialzie a list of coordinates that will be ordered such that the first entry in the list is the top-left,
+ the second entry is the top-right, the third is the bottom-right, and the fourth is the bottom-left
+ :param pts: array containing 4 points
+ :return: ordered list of 4 points
+ """
rect = np.zeros((4, 2), dtype="float32")
# the top-left point will have the smallest sum, whereas
@@ -92,8 +157,14 @@
return rect
-# www.pyimagesearch.com/2014/08/25/4-point-opencv-getperspective-transform-example/
def four_point_transform(image, pts):
+ """
+ Transform a quadrilateral section of an image into a rectangular area
+ From: www.pyimagesearch.com/2014/08/25/4-point-opencv-getperspective-transform-example/
+ :param image: source image
+ :param pts: 4 corners of the quadrilateral
+ :return: rectangular image of the specified area
+ """
# obtain a consistent order of the points and unpack them
# individually
rect = order_points(pts)
@@ -138,71 +209,14 @@
return warped
-# Get the names of the output layers
-def get_outputs_names(net):
- # Get the names of all the layers in the network
- layers_names = net.getLayerNames()
- # Get the names of the output layers, i.e. the layers with unconnected outputs
- return [layers_names[i[0] - 1] for i in net.getUnconnectedOutLayers()]
-
-
-# Remove the bounding boxes with low confidence using non-maxima suppression
-# https://www.learnopencv.com/deep-learning-based-object-detection-using-yolov3-with-opencv-python-c/
-def post_process(frame, outs, thresh_conf, thresh_nms):
- frame_height = frame.shape[0]
- frame_width = frame.shape[1]
-
- # Scan through all the bounding boxes output from the network and keep only the
- # ones with high confidence scores. Assign the box's class label as the class with the highest score.
- class_ids = []
- confidences = []
- boxes = []
- for out in outs:
- for detection in out:
- scores = detection[5:]
- class_id = np.argmax(scores)
- confidence = scores[class_id]
- if confidence > thresh_conf:
- center_x = int(detection[0] * frame_width)
- center_y = int(detection[1] * frame_height)
- width = int(detection[2] * frame_width)
- height = int(detection[3] * frame_height)
- left = int(center_x - width / 2)
- top = int(center_y - height / 2)
- class_ids.append(class_id)
- confidences.append(float(confidence))
- boxes.append([left, top, width, height])
-
- # Perform non maximum suppression to eliminate redundant overlapping boxes with lower confidences.
- indices = [ind[0] for ind in cv2.dnn.NMSBoxes(boxes, confidences, thresh_conf, thresh_nms)]
-
- ret = [[class_ids[i], confidences[i], boxes[i]] for i in indices]
- return ret
-
-
-# Draw the predicted bounding box
-def draw_pred(frame, class_id, classes, conf, left, top, right, bottom):
- # Draw a bounding box.
- cv2.rectangle(frame, (left, top), (right, bottom), (0, 0, 255))
-
- label = '%.2f' % conf
-
- # Get the label for the class name and its confidence
- if classes:
- assert (class_id < len(classes))
- label = '%s:%s' % (classes[class_id], label)
-
- # Display the label at the top of the bounding box
- label_size, base_line = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
- top = max(top, label_size[1])
- cv2.putText(frame, label, (left, top), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255))
-
-
def remove_glare(img):
"""
+ Reduce the effect of glaring in the image
Inspired from:
http://www.amphident.de/en/blog/preprocessing-for-automatic-pattern-identification-in-wildlife-removing-glare.html
The idea is to find area that has low saturation but high value, which is what a glare usually look like.
+ :param img: source image
+ :return: corrected image with glaring smoothened out
"""
img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
_, s, v = cv2.split(img_hsv)
@@ -225,66 +239,148 @@
return corrected
-def find_card(img, thresh_c=5, kernel_size=(3, 3), size_ratio=0.2):
+def find_card(img, thresh_c=5, kernel_size=(3, 3), size_thresh=10000, debug=False):
+ """
+ Find contours of all cards in the image
+ :param img: source image
+ :param thresh_c: value of the constant C for adaptive thresholding
+ :param kernel_size: dimension of the kernel used for dilation and erosion
+ :param size_thresh: threshold for size (in pixel) of the contour to be a candidate
+ :return: list of candidate contours
+ """
# Typical pre-processing - grayscale, blurring, thresholding
img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img_blur = cv2.medianBlur(img_gray, 5)
- img_thresh = cv2.adaptiveThreshold(img_blur, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY_INV, 5, thresh_c)
-
+ img_thresh = cv2.adaptiveThreshold(img_blur, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY_INV, 11, thresh_c)
+ if debug:
+ cv2.imshow('Thres', img_thresh)
# Dilute the image, then erode them to remove minor noises
kernel = np.ones(kernel_size, np.uint8)
img_dilate = cv2.dilate(img_thresh, kernel, iterations=1)
img_erode = cv2.erode(img_dilate, kernel, iterations=1)
-
+ if debug:
+ cv2.imshow('Eroded', img_erode)
# Find the contour
- #img_contour = img_erode.copy()
- _, cnts, hier = cv2.findContours(img_erode, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
+ cnts, hier = cv2.findContours(img_erode, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
if len(cnts) == 0:
print('no contours')
return []
- #img_contour = cv2.cvtColor(img_contour, cv2.COLOR_GRAY2BGR)
- #img_contour = cv2.drawContours(img_contour, cnts, -1, (0, 255, 0), 1)
- #cv2.imshow('test', img_contour)
-
- # For each contours detected, check if they are large enough and are rectangle
+ img_cont = cv2.cvtColor(img_erode, cv2.COLOR_GRAY2BGR)
+ img_cont_base = img_cont.copy()
+ cnts2 = sorted(cnts, key=cv2.contourArea, reverse=True)
+ cnts2 = cnts2[:10]
+ for i in range(0, len(cnts2)):
+ print(i, len(cnts2[i]))
+ if debug:
+ cv2.drawContours(img_cont, cnts2, -1, (0, 255, 0), 3)
+ cv2.imshow('Contours', img_cont)
+ # The hierarchy from cv2.findContours() is similar to a tree: each node has an access to the parent, the first child
+ # their previous and next node
+ # Using recursive search, find the uppermost contour in the hierarchy that satisfies the condition
+ # The candidate contour must be rectangle (has 4 points) and should be larger than a threshold
cnts_rect = []
- ind_sort = sorted(range(len(cnts)), key=lambda i: cv2.contourArea(cnts[i]), reverse=True)
- for i in range(min(len(cnts), 5)): # The card should be within top 5 largest contour
- size = cv2.contourArea(cnts[ind_sort[i]])
- peri = cv2.arcLength(cnts[ind_sort[i]], True)
- approx = cv2.approxPolyDP(cnts[ind_sort[i]], 0.04 * peri, True)
- if size > img.shape[0] * img.shape[1] * size_ratio and len(approx) == 4:
- cnts_rect.append(approx)
+ stack = [(0, hier[0][0])]
+ while len(stack) > 0:
+ i_cnt, h = stack.pop()
+ i_next, i_prev, i_child, i_parent = h
+ if i_next != -1:
+ stack.append((i_next, hier[0][i_next]))
+ cnt = cnts[i_cnt]
+ size = cv2.contourArea(cnt)
+ peri = cv2.arcLength(cnt, True)
+ approx = cv2.approxPolyDP(cnt, 0.04 * peri, True)
+ print('Base Size:', size)
+ print('Len Approx:', len(approx))
+ if size >= size_thresh and len(approx) == 4:
+ # lets see if we got a contour very close in size as child
+ if i_child != -1:
+ img_ccont = img_cont_base.copy()
+ # lets collect all children
+ c_list = [cnts[i_child]]
+ h_info = hier[0][i_child]
+ while h_info[0] != -1:
+ cld = cnts[h_info[0]]
+ c_list.append(cld)
+ h_info = hier[0][h_info[0]]
+ # child with biggest area
+ c_list.sort(key=cv2.contourArea, reverse=True)
+ c_cnt = c_list[0] # the biggest child
+ if debug:
+ cv2.drawContours(img_ccont, c_list[:1], -1, (0, 255, 0), 1)
+ cv2.imshow('CCont', img_ccont)
+ c_size = cv2.contourArea(c_cnt)
+ c_approx = cv2.approxPolyDP(c_cnt, 0.04 * peri, True)
+ if len(c_approx) == 4 and (c_size/size) > 0.85:
+ rect = cv2.minAreaRect(c_cnt)
+ box = cv2.boxPoints(rect)
+ box = np.intp(box)
+ print(c_cnt)
+ print(box)
+ print('CSize:', c_size, '%:', c_size/size)
+ b2 = []
+ for x in box:
+ b2.append([x])
+ cnts_rect.append(np.array(b2))
+ else:
+ print('CF:', (c_size/size))
+ print('Size:', size)
+ cnts_rect.append(approx)
+ else:
+ #print('CF:', (c_size/size))
+ print('Size:', size)
+ cnts_rect.append(approx)
+ else:
+ if i_child != -1:
+ stack.append((i_child, hier[0][i_child]))
return cnts_rect
def draw_card_graph(exist_cards, card_pool, f_len):
- w_card = 63
+ """
+ Given the history of detected cards in the current and several previous frames, draw a simple graph
+ displaying the detected cards with its confidence level
+ :param exist_cards: History of all detected cards in the previous (f_len) frames
+ :param card_pool: pandas dataframe of all card's information
+ :param f_len: length of windows (in frames) to consider for confidence level
+ :return:
+ """
+ # Lots of constants to set the dimension of each elements
+ w_card = 63 # Width of the card image displayed
h_card = 88
- gap = 25
- gap_sm = 10
- w_bar = 300
+ gap = 25 # Offset between each elements
+ gap_sm = 10 # Small offset
+ w_bar = 300 # Length of the confidence bar at 100%
h_bar = 12
txt_scale = 0.8
- n_cards_p_col = 4
- w_img = gap + (w_card + gap + w_bar + gap) * 2
- #h_img = gap + (h_card + gap) * n_cards_p_col
+ n_cards_p_col = 4 # Number of cards displayed per one column
+ w_img = gap + (w_card + gap + w_bar + gap) * 2 # Dimension of the entire graph (for 2 columns)
h_img = 480
img_graph = np.zeros((h_img, w_img, 3), dtype=np.uint8)
x_anchor = gap
y_anchor = gap
i = 0
+
+ # Cards are displayed from the most confident to the least
+ # Confidence level is calculated by number of frames that the card was detected in
for key, val in sorted(exist_cards.items(), key=itemgetter(1), reverse=True)[:n_cards_p_col * 2]:
card_name = key[:key.find('(') - 1]
card_set = key[key.find('(') + 1:key.find(')')]
confidence = sum(val) / f_len
card_info = card_pool[(card_pool['name'] == card_name) & (card_pool['set'] == card_set)].iloc[0]
- img_name = '%s/card_img/tiny/%s/%s_%s.png' % (transform_data.data_dir, card_info['set'],
- card_info['collector_number'],
- fetch_data.get_valid_filename(card_info['name']))
- card_img = cv2.imread(img_name)
+ img_name = '%s/card_img/tiny/%s/%s_%s.png' % (Config.data_dir, card_info['set'],
+ card_info['collector_number'],
+ fetch_data.get_valid_filename(card_info['name']))
+ # If the card image is not found, just leave it blank
+ if os.path.exists(img_name):
+ card_img = cv2.imread(img_name)
+ else:
+ card_img = np.ones((h_card, w_card, 3)) * 255
+ cv2.putText(card_img, 'X', ((w_card - int(txt_scale * 25)) // 2, (h_card + int(txt_scale * 25)) // 2),
+ cv2.FONT_HERSHEY_SIMPLEX, txt_scale, (0, 0, 0), 2)
+
+ # Insert the card image, card name, and confidence bar to the graph
img_graph[y_anchor:y_anchor + h_card, x_anchor:x_anchor + w_card] = card_img
cv2.putText(img_graph, '%s (%s)' % (card_name, card_set),
(x_anchor + w_card + gap, y_anchor + gap_sm + int(txt_scale * 25)), cv2.FONT_HERSHEY_SIMPLEX,
@@ -301,266 +397,346 @@
return img_graph
-def detect_frame(net, classes, img, card_pool, thresh_conf=0.5, thresh_nms=0.4, in_dim=(416, 416), out_path=None, display=True,
- debug=False):
- start_1 = time.time()
- elapsed = []
- '''
- # Create a 4D blob from a frame.
- blob = cv2.dnn.blobFromImage(img, 1 / 255, in_dim, [0, 0, 0], 1, crop=False)
+def detect_frame(img, card_pool, hash_size=32, size_thresh=10000,
+ out_path=None, display=True, debug=False):
+ """
+ Identify all cards in the input frame, display or save the frame if needed
+ :param img: input frame
+ :param card_pool: pandas dataframe of all card's information
+ :param hash_size: param for pHash algorithm
+ :param size_thresh: threshold for size (in pixel) of the contour to be a candidate
+ :param out_path: path to save the result
+ :param display: flag for displaying the result
+ :param debug: flag for debug mode
+ :return: list of detected card's name/set and resulting image
+ """
- # Sets the input to the network
- net.setInput(blob)
-
- # Runs the forward pass to get output of the output layers
- outs = net.forward(get_outputs_names(net))
- elapsed.append((time.time() - start_1) * 1000)
-
- start_2 = time.time()
- img_result = img.copy()
-
- # Remove the bounding boxes with low confidence
- obj_list = post_process(img, outs, thresh_conf, thresh_nms)
- for obj in obj_list:
- class_id, confidence, box = obj
- left, top, width, height = box
- draw_pred(img_result, class_id, classes, confidence, left, top, left + width, top + height)
- elapsed.append((time.time() - start_2) * 1000)
- '''
- img_result = img.copy()
- obj_list = []
- # Put efficiency information. The function getPerfProfile returns the
- # overall time for inference(t) and the timings for each of the layers(in layersTimes)
- #if display:
- # t, _ = net.getPerfProfile()
- # label = 'Inference time: %.2f ms' % (t * 1000.0 / cv2.getTickFrequency())
- # cv2.putText(img_result, label, (0, 15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255))
-
- '''
- Assuming that the model has properly identified all cards, there should be 1 card that can be classified per
- bounding box. Find the largest rectangular contour from the region of interest, and identify the card by
- comparing the perceptual hashing of the image with the other cards' image from the database.
- '''
+ img_result = img.copy() # For displaying and saving
det_cards = []
- for i in range(len(obj_list)):
- start_3 = time.time()
- _, _, box = obj_list[i]
- left, top, width, height = box
- # Just in case the bounding box trimmed the edge of the cards, give it a bit of offset around the edge
- offset_ratio = 0.1
- x1 = max(0, int(left - offset_ratio * width))
- x2 = min(img.shape[1], int(left + (1 + offset_ratio) * width))
- y1 = max(0, int(top - offset_ratio * height))
- y2 = min(img.shape[0], int(top + (1 + offset_ratio) * height))
- img_snip = img[y1:y2, x1:x2]
- cnts = find_card(img_snip)
- elapsed.append((time.time() - start_3) * 1000)
- if len(cnts) > 0:
- start_4 = time.time()
- cnt = cnts[0] # The largest (rectangular) contour
- pts = np.float32([p[0] for p in cnt])
- img_warp = four_point_transform(img_snip, pts)
- img_warp = cv2.resize(img_warp, (card_width, card_height))
- elapsed.append((time.time() - start_4) * 1000)
- '''
- img_art = img_warp[47:249, 22:294]
- img_art = Image.fromarray(img_art.astype('uint8'), 'RGB')
- art_hash = ih.phash(img_art, hash_size=32, highfreq_factor=4)
- card_pool['hash_diff'] = card_pool['art_hash'] - art_hash
- min_cards = card_pool[card_pool['hash_diff'] == min(card_pool['hash_diff'])]
- card_name = min_cards.iloc[0]['name']
- '''
- start_5 = time.time()
- img_card = Image.fromarray(img_warp.astype('uint8'), 'RGB')
- card_hash = ih.phash(img_card, hash_size=32, highfreq_factor=4).hash.flatten()
- card_pool['hash_diff'] = card_pool['card_hash'].apply(lambda x: np.count_nonzero(x != card_hash))
- min_cards = card_pool[card_pool['hash_diff'] == min(card_pool['hash_diff'])]
- card_name = min_cards.iloc[0]['name']
- card_set = min_cards.iloc[0]['set']
- det_cards.append((card_name, card_set))
- hash_diff = min_cards.iloc[0]['hash_diff']
- elapsed.append((time.time() - start_5) * 1000)
+ # Detect contours of all cards in the image
+ cnts = find_card(img_result, size_thresh=size_thresh, debug=debug)
+ print('Countours:', len(cnts))
+ for i in range(len(cnts)):
+ print('Contour', i)
+ cnt = cnts[i]
+ # For the region of the image covered by the contour, transform them into a rectangular image
+ pts = np.float32([p[0] for p in cnt])
+ img_warp = four_point_transform(img, pts)
- # Display the result
- if debug:
- # cv2.rectangle(img_warp, (22, 47), (294, 249), (0, 255, 0), 2)
- cv2.putText(img_warp, card_name + ', ' + str(hash_diff), (0, 50),
- cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2)
- cv2.putText(img_result, card_name, (x1, y1), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 2)
- if debug:
- cv2.imshow('card#%d' % i, img_warp)
- elif debug:
- cv2.imshow('card#%d' % i, np.zeros((1, 1), dtype=np.uint8))
+ # To identify the card from the card image, perceptual hashing (pHash) algorithm is used
+ # Perceptual hash is a hash string built from features of the input medium. If two media are similar
+ # (ie. has similar features), their resulting pHash value will be very close.
+ # Using this property, the matching card for the given card image can be found by comparing pHash of
+ # all cards in the database, then finding the card that results in the minimal difference in pHash value.
+ '''
+ img_art = img_warp[47:249, 22:294]
+ img_art = Image.fromarray(img_art.astype('uint8'), 'RGB')
+ art_hash = ih.phash(img_art, hash_size=hash_size).hash.flatten()
+ card_pool['hash_diff'] = card_pool['art_hash'].apply(lambda x: np.count_nonzero(x != art_hash))
+ '''
+ img_card = Image.fromarray(img_warp.astype('uint8'), 'RGB')
+ img_card_size = img_warp.shape
+ print(img_card_size)
+ cut = [round(img_card_size[0]*0.57),round(img_card_size[0]*0.615),round(img_card_size[1]*0.81),round(img_card_size[1]*0.940)]
+ print(cut)
+ img_set_part = img_warp[cut[0]:cut[1], cut[2]:cut[3]]
+ print(img_set_part.shape)
+ img_set = Image.fromarray(img_set_part.astype('uint8'), 'RGB')
+ print('img set')
+ if debug:
+ cv2.imshow("Set Img#%d" % i, img_set_part)
+
+ # the stored values of hashes in the dataframe is pre-emptively flattened already to minimize computation time
+ card_hash = ih.phash(img_card, hash_size=hash_size).hash.flatten()
+ card_pool['hash_diff'] = card_pool['card_hash_%d' % hash_size]
+ card_pool['hash_diff'] = card_pool['hash_diff'].apply(lambda x: np.count_nonzero(x != card_hash))
+ min_card = card_pool[card_pool['hash_diff'] == min(card_pool['hash_diff'])].iloc[0]
+ hash_diff = min_card['hash_diff']
+
+ top_matches = sorted(card_pool['hash_diff'])
+ card_one = card_pool[card_pool['hash_diff'] == top_matches[0]].iloc[0]
+ card_two = card_pool[card_pool['hash_diff'] == top_matches[1]].iloc[0]
+
+ if card_one['name'] == card_two['name'] and card_one['set'] != card_two['set']:
+ set_img_hash = ih.whash(img_set, hash_size=hash_size).hash.flatten()
+ cd_data = pd.DataFrame(columns=list(card_pool.columns.values))
+ print(list(card_pool.columns.values))
+ candidates = []
+ for ix in range(0, 2):
+ cd = card_pool[card_pool['hash_diff'] == top_matches[ix]].iloc[0]
+ cd_data.loc[0 if cd_data.empty else cd_data.index.max()+1] = cd
+ print('Idx:', ix, 'Name:', cd['name'], 'Set:', cd['set'], 'Diff:', top_matches[ix])
+
+
+ cd_data['set_hash_diff'] = cd_data['set_hash_%d' % 64]
+ cd_data['set_hash_diff'] = cd_data['set_hash_diff'].apply(lambda x: np.count_nonzero(x != set_img_hash))
+ conf = sorted(cd_data['set_hash_diff'])
+ print('Confs:', conf)
+ best_match = cd_data[cd_data['set_hash_diff'] == min(cd_data['set_hash_diff'])].iloc[0]
+ print('Best Match', 'Name:', best_match['name'], 'Set:', best_match['set'])
+
+ min_card = best_match
+ card_name = min_card['name']
+ card_set = min_card['set']
+ det_cards.append((card_name, card_set))
+
+ # Render the result, and display them if needed
+ cv2.drawContours(img_result, [cnt], -1, (0, 255, 0), 2)
+ cv2.putText(img_result, card_name, (min(pts[0][0], pts[1][0]), min(pts[0][1], pts[1][1])),
+ cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 2)
+ if debug:
+ # cv2.rectangle(img_warp, (22, 47), (294, 249), (0, 255, 0), 2)
+ cv2.putText(img_warp, card_name + ':' + card_set + ', ' + str(hash_diff), (0, 20),
+ cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
+ cv2.imshow('card#%d' % i, img_warp)
+ if display:
+ cv2.imshow('Result', img_result)
+ inp = cv2.waitKey(0)
if out_path is not None:
+ print(out_path)
cv2.imwrite(out_path, img_result.astype(np.uint8))
- elapsed = [(time.time() - start_1) * 1000] + elapsed
- #print(', '.join(['%.2f' % t for t in elapsed]))
- return obj_list, det_cards, img_result
+ return det_cards, img_result
-def detect_video(net, classes, capture, card_pool, thresh_conf=0.5, thresh_nms=0.4, in_dim=(416, 416), out_path=None,
- display=True, debug=False):
+def detect_video(capture, card_pool, hash_size=32, size_thresh=10000,
+ out_path=None, display=True, show_graph=True, debug=False, crop_x=0, crop_y=0):
+ """
+ Identify all cards in the continuous video stream, display or save the result if needed
+ :param capture: input video stream
+ :param card_pool: pandas dataframe of all card's information
+ :param hash_size: param for pHash algorithm
+ :param size_thresh: threshold for size (in pixel) of the contour to be a candidate
+ :param out_path: path to save the result
+ :param display: flag for displaying the result
+ :param show_graph: flag to show graph
+ :param debug: flag for debug mode
+ :return: list of detected card's name/set and resulting image
+ :return:
+ """
+ # Get the dimension of the output video, and set it up
+ if show_graph:
+ img_graph = draw_card_graph({}, pd.DataFrame(), -1) # Black image of the graph just to get the dimension
+ width = round(capture.get(cv2.CAP_PROP_FRAME_WIDTH)) - 2*crop_x + img_graph.shape[1]
+ height = max(round(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)) - 2*crop_y, img_graph.shape[0])
+ height += 200 # some space to display last detected cards
+ else:
+ width = round(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+ height = round(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
if out_path is not None:
- img_graph = draw_card_graph({}, None, -1) # Black image of the graph just to get the dimension
- width = round(capture.get(cv2.CAP_PROP_FRAME_WIDTH)) + img_graph.shape[1]
- height = max(round(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)), img_graph.shape[0])
vid_writer = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*'MJPG'), 10.0, (width, height))
max_num_obj = 0
f_len = 10 # number of frames to consider to check for existing cards
exist_cards = {}
+
+ exist_card_single = {}
+ written_out_cards = set()
+ found_cards = []
try:
while True:
ret, frame = capture.read()
+ croped_img = frame[crop_y:-crop_y, crop_x:-crop_x]
+ fimg = cv2.flip(croped_img, -1)
start_time = time.time()
if not ret:
# End of video
print("End of video. Press any key to exit")
cv2.waitKey(0)
break
- # Use the YOLO model to identify each cards annonymously
- start_yolo = time.time()
- obj_list, det_cards, img_result = detect_frame(net, classes, frame, card_pool, thresh_conf=thresh_conf,
- thresh_nms=thresh_nms, in_dim=in_dim, out_path=None,
- display=display, debug=debug)
- elapsed_yolo = (time.time() - start_yolo) * 1000
- # If the card was already detected in the previous frame, append 1 to the list
- # If the card previously detected was not found in this trame, append 0 to the list
- # If the card wasn't previously detected, make a new list and add 1 to it
- # If the same card is detected multiple times in the same frame, keep track of the duplicates
- # The confidence will be calculated based on the number of frames the card was detected for
- det_cards_count = collections.Counter(det_cards).items()
- det_cards_list = []
- for card, count in det_cards_count:
- card_name, card_set = card
- for i in range(count): 1
- key = '%s (%s) #%d' % (card_name, card_set, i + 1)
- det_cards_list.append(key)
- gone = []
- for key, val in exist_cards.items():
- if key in det_cards_list:
- exist_cards[key] = exist_cards[key][1 - f_len:] + [1]
- else:
- exist_cards[key] = exist_cards[key][1 - f_len:] + [0]
- if len(val) == f_len and sum(val) == 0:
- gone.append(key)
- for key in det_cards_list:
- if key not in exist_cards.keys():
- exist_cards[key] = [1]
- for key in gone:
- exist_cards.pop(key)
- start_graph = time.time()
- img_graph = draw_card_graph(exist_cards, card_pool, f_len)
- elapsed_graph = (time.time() - start_graph) * 1000
- if debug:
- max_num_obj = max(max_num_obj, len(obj_list))
- for i in range(len(obj_list), max_num_obj):
- cv2.imshow('card#%d' % i, np.zeros((1, 1), dtype=np.uint8))
+ # Detect all cards from the current frame
+ det_cards, img_result = detect_frame(fimg, card_pool, hash_size=hash_size, size_thresh=size_thresh,
+ out_path=None, display=False, debug=debug)
+ if show_graph:
+ # If the card was already detected in the previous frame, append 1 to the list
+ # If the card previously detected was not found in this trame, append 0 to the list
+ # If the card wasn't previously detected, make a new list and add 1 to it
+ # If the same card is detected multiple times in the same frame, keep track of the duplicates
+ # The confidence will be calculated based on the number of frames the card was detected for
+ det_cards_count = collections.Counter(det_cards).items()
+ det_cards_list = []
+ for card, count in det_cards_count:
+ card_name, card_set = card
+ for i in range(count): 1
+ key = '%s (%s) #%d' % (card_name, card_set, i + 1)
+ det_cards_list.append(key)
+ gone = []
+ for key, val in exist_cards.items():
+ if key in det_cards_list:
+ exist_cards[key] = exist_cards[key][1 - f_len:] + [1]
+ else:
+ exist_cards[key] = exist_cards[key][1 - f_len:] + [0]
+ if len(val) == f_len and sum(val) == 0:
+ gone.append(key) # not there anymore
- start_display = time.time()
- img_save = np.zeros((height, width, 3), dtype=np.uint8)
- img_save[0:img_result.shape[0], 0:img_result.shape[1]] = img_result
- img_save[0:img_graph.shape[0], img_result.shape[1]:img_result.shape[1] + img_graph.shape[1]] = img_graph
+ det_card_map = {}
+ gone_single = []
+ for card_name, card_set in det_cards:
+ skey = '%s (%s)' % (card_name, card_set)
+ det_card_map[skey] = (card_name, card_set)
+
+ for key, val in exist_card_single.items():
+ if key in det_card_map:
+ exist_card_single[key] = exist_card_single[key][1 - f_len:] + [1]
+ else:
+ exist_card_single[key] = exist_card_single[key][1 - f_len:] + [0]
+
+ if len(val) == f_len and sum(val) == 0:
+ gone_single.append(key)
+ if key in written_out_cards:
+ written_out_cards.remove(key)
+ if len(val) == f_len and sum(val) == f_len:
+ if key not in written_out_cards and key in det_card_map:
+ written_out_cards.add(key)
+ found_cards.append(det_card_map[key])
+
+ for key in det_card_map:
+ if key not in exist_card_single.keys():
+ exist_card_single[key] = [1]
+ for key in gone_single:
+ exist_card_single.pop(key)
+
+
+ for key in det_cards_list:
+ if key not in exist_cards.keys():
+ exist_cards[key] = [1]
+ for key in gone:
+ exist_cards.pop(key)
+
+
+ # Draw the graph based on the history of detected cards, then concatenate it with the result image
+ img_graph = draw_card_graph(exist_cards, card_pool, f_len)
+ img_save = np.zeros((height, width, 3), dtype=np.uint8)
+ img_save[0:img_result.shape[0], 0:img_result.shape[1]] = img_result
+ img_save[0:img_graph.shape[0], img_result.shape[1]:img_result.shape[1] + img_graph.shape[1]] = img_graph
+ for c, card in enumerate(reversed(found_cards[-10:]), 1):
+ cv2.putText(img_save, f'{card[0]} ({card[1].upper()})',(0, height-200+18*c), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0))
+ else:
+ img_save = img_result
+
+ # Display the result
if display:
cv2.imshow('result', img_save)
- elapsed_display = (time.time() - start_display) * 1000
+ if debug:
+ max_num_obj = max(max_num_obj, len(det_cards))
+ for i in range(len(det_cards), max_num_obj):
+ cv2.imshow('card#%d' % i, np.zeros((1, 1), dtype=np.uint8))
elapsed_ms = (time.time() - start_time) * 1000
- #print('Elapsed time: %.2f ms, %.2f, %.2f, %.2f' % (elapsed_ms, elapsed_yolo, elapsed_graph, elapsed_display))
+ print('Elapsed time: %.2f ms' % elapsed_ms)
if out_path is not None:
vid_writer.write(img_save.astype(np.uint8))
- cv2.waitKey(1)
+ inp = cv2.waitKey(0)
+ if 'q' == chr(inp & 255):
+ break
except KeyboardInterrupt:
capture.release()
if out_path is not None:
vid_writer.release()
cv2.destroyAllWindows()
+ with open('detect.txt', 'w') as of:
+ counter = collections.Counter(found_cards)
+ for key in counter:
+ of.write(f'{counter[key]} [{key[1].upper()}] {key[0]}\n')
-def main():
+
+
+def main(args):
# Specify paths for all necessary files
- test_path = os.path.abspath('test_file/test4.mp4')
- #weight_path = 'backup/tiny_yolo_10_39500.weights'
- #cfg_path = 'cfg/tiny_yolo_10.cfg'
- #class_path = "data/obj_10.names"
- weight_path = 'weights/second_general/tiny_yolo_final.weights'
- cfg_path = 'cfg/tiny_yolo_old.cfg'
- class_path = 'data/obj.names'
- out_dir = 'out'
- if not os.path.isfile(test_path):
- print('The test file %s doesn\'t exist!' % os.path.abspath(test_path))
- return
- if not os.path.isfile(weight_path):
- print('The weight file %s doesn\'t exist!' % os.path.abspath(test_path))
- return
- if not os.path.isfile(cfg_path):
- print('The config file %s doesn\'t exist!' % os.path.abspath(test_path))
- return
- if not os.path.isfile(class_path):
- print('The class file %s doesn\'t exist!' % os.path.abspath(test_path))
- return
+ hash_sizes = {16, 32}
+ hash_sizes.add(args.hash_size)
+ pck_path = os.path.abspath('card_pool.pck')
+ if os.path.isfile(pck_path):
+ card_pool = pd.read_pickle(pck_path)
+ else:
+ print('Warning: pickle for card database %s is not found!' % pck_path)
+ # Merge database for all cards, then calculate pHash values of each, store them
+ df_list = []
+ for set_name in Config.all_set_list:
+ if set_name == 'con':
+ set_name = 'con__'
+ csv_name = '%s/csv/%s.csv' % (Config.data_dir, set_name)
+ df = fetch_data.load_all_cards_text(csv_name)
+ df_list.append(df)
+ card_pool = pd.concat(df_list, sort=True)
+ card_pool.reset_index(drop=True, inplace=True)
+ card_pool.drop('Unnamed: 0', axis=1, inplace=True, errors='ignore')
+ card_pool = calc_image_hashes(card_pool, save_to=pck_path, hash_size=hash_sizes)
+ ch_key = 'card_hash_%d' % args.hash_size
+ set_key = 'set_hash_%d' % 64
+ if ch_key not in card_pool.columns:
+ # we did not generate this hash_size yet
+ print('We need to add hash_size=%d' % (args.hash_size,))
+ card_pool = calc_image_hashes(card_pool, save_to=pck_path, hash_size=[args.hash_size])
+ card_pool = card_pool[['name', 'set', 'collector_number', ch_key, set_key]]
- '''
- df_list = []
- for set_name in fetch_data.all_set_list:
- csv_name = '%s/csv/%s.csv' % (transform_data.data_dir, set_name)
- df = fetch_data.load_all_cards_text(csv_name)
- df_list.append(df)
- #print(df)
- card_pool = pd.concat(df_list, sort=True)
- card_pool.reset_index(drop=True, inplace=True)
- card_pool.drop('Unnamed: 0', axis=1, inplace=True, errors='ignore')
- for hash_size in [8, 16, 32, 64]:
- for highfreq_factor in [4, 8, 16, 32]:
- pck_name = 'card_pool_%d_%d.pck' % (hash_size, highfreq_factor)
- if not os.path.exists(pck_name):
- print(pck_name)
- calc_image_hashes(card_pool, save_to=pck_name, hash_size=hash_size, highfreq_factor=highfreq_factor)
- '''
- #csv_name = '%s/csv/%s.csv' % (transform_data.data_dir, 'rtr')
- #card_pool = fetch_data.load_all_cards_text(csv_name)
- #card_pool = calc_image_hashes(card_pool, save_to='card_pool.pck')
- #return
- card_pool = pd.read_pickle('card_pool_32_4.pck')
- #card_pool = card_pool[(card_pool['set'] == 'rtr') | (card_pool['set'] == 'isd')]
- card_pool = card_pool[['name', 'set', 'collector_number', 'card_hash']]
+ # Processing time is almost linear to the size of the database
+ # Program can be much faster if the search scope for the card can be reduced
+ card_pool = card_pool[card_pool['set'].isin(Config.set_2003_list)]
# ImageHash is basically just one numpy.ndarray with (hash_size)^2 number of bits. pre-emptively flattening it
# significantly increases speed for subtracting hashes in the future.
- card_pool['card_hash'] = card_pool['card_hash'].apply(lambda x: x.hash.flatten())
+ card_pool[ch_key] = card_pool[ch_key].apply(lambda x: x.hash.flatten())
+ card_pool[set_key] = card_pool[set_key].apply(lambda x: x.hash.flatten())
- thresh_conf = 0.01
- thresh_nms = 0.8
+ # If the test file isn't given, use webcam to capture video
+ if args.in_path is None:
+ capture = cv2.VideoCapture(0, cv2.CAP_V4L)
+ capture.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*"MJPG"))
+ capture.set(cv2.CAP_PROP_FRAME_WIDTH, 1920)
+ capture.set(cv2.CAP_PROP_FRAME_HEIGHT, 1080)
- # Setup
- # Read class names from text file
- with open(class_path, 'r') as f:
- classes = [line.strip() for line in f.readlines()]
- # Load up the neural net using the config and weights
- net = cv2.dnn.readNetFromDarknet(cfg_path, weight_path)
- net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
- net.setPreferableTarget(cv2.dnn.DNN_TARGET_CPU)
-
- # Save the detection result if out_dir is provided
- if out_dir is None or out_dir == '':
- out_path = None
- else:
- f_name = os.path.split(test_path)[1]
- out_path = out_dir + '/' + f_name[:f_name.find('.')] + '.avi'
- # Check if test file is image or video
- test_ext = test_path[test_path.find('.') + 1:]
-
- if test_ext in ['jpg', 'jpeg', 'bmp', 'png', 'tiff']:
- img = cv2.imread(test_path)
- detect_frame(net, classes, img, card_pool, out_path=out_path, thresh_conf=thresh_conf, thresh_nms=thresh_nms)
- else:
- capture = cv2.VideoCapture(0)
- detect_video(net, classes, capture, card_pool, out_path=out_path, thresh_conf=thresh_conf, thresh_nms=thresh_nms,
- display=True, debug=False)
+ thres = int(((1920-2*500)*(1080-2*200)*0.3))
+ print('Threshold:', thres)
+ detect_video(capture, card_pool, hash_size=args.hash_size, out_path='%s/result.avi' % args.out_path,
+ display=args.display, show_graph=args.show_graph, debug=args.debug, crop_x=500, crop_y=200, size_thresh=thres)
capture.release()
+ else:
+ # Save the detection result if args.out_path is provided
+ if args.out_path is None:
+ out_path = None
+ else:
+ f_name = os.path.split(args.in_path)[1]
+ out_path = '%s/%s.avi' % (args.out_path, f_name[:f_name.find('.')])
+
+ if not os.path.isfile(args.in_path):
+ print('The test file %s doesn\'t exist!' % os.path.abspath(args.in_path))
+ return
+ # Check if test file is image or video
+ test_ext = args.in_path[args.in_path.find('.') + 1:]
+ if test_ext in ['jpg', 'jpeg', 'bmp', 'png', 'tiff']:
+ # Test file is an image
+ img = cv2.imread(args.in_path)
+ if img is None:
+ print('Could not read', args.in_path)
+ detect_frame(img, card_pool, hash_size=args.hash_size, out_path=out_path, display=args.display,
+ debug=args.debug)
+ else:
+ # Test file is a video
+ capture = cv2.VideoCapture(args.in_path)
+ detect_video(capture, card_pool, hash_size=args.hash_size, out_path=out_path, display=args.display,
+ show_graph=args.show_graph, debug=args.debug)
+ capture.release()
pass
if __name__ == '__main__':
- main()
+ parser = argparse.ArgumentParser()
+ parser.add_argument('-i', '--in', dest='in_path', help='Path of the input file. For webcam, leave it blank',
+ type=str)
+ parser.add_argument('-o', '--out', dest='out_path', help='Path of the output directory to save the result',
+ type=str)
+ parser.add_argument('-hs', '--hash_size', dest='hash_size',
+ help='Size of the hash for pHash algorithm', type=int, default=16)
+ parser.add_argument('-dsp', '--display', dest='display', help='Display the result', action='store_true',
+ default=False)
+ parser.add_argument('-dbg', '--debug', dest='debug', help='Enable debug mode', action='store_true', default=False)
+ parser.add_argument('-gph', '--show_graph', dest='show_graph', help='Display the graph for video output',
+ action='store_true', default=False)
+ args = parser.parse_args()
+ if not args.display and args.out_path is None:
+ # Then why the heck are you running this thing in the first place?
+ print('The program isn\'t displaying nor saving any output file. Please change the setting and try again.')
+ exit()
+ main(args)
--
Gitblit v1.10.0