added some code related to finding the set

SpeedProg

2019-08-23 ff863fe7f8540a10e699e445317d6b2399c51440

 opencv_dnn.py

@@ -9,7 +9,7 @@
import pandas as pd
from PIL import Image
import time

from multiprocessing import Pool
from config import Config
import fetch_data

@@ -22,25 +22,14 @@
https://github.com/hj3yoo/mtg_card_detector/tree/dea64611730c84a59c711c61f7f80948f82bcd31 
"""


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]
    
    # Since some double-faced cards may result in two different cards, create a new dataframe to store the result
def do_calc(args):
    card_pool = args[0]
    hash_size = args[1]
    new_pool = pd.DataFrame(columns=list(card_pool.columns.values))
    for hs in hash_size:
            new_pool['card_hash_%d' % hs] = np.NaN
            #new_pool['art_hash_%d' % hs] = np.NaN
        new_pool['card_hash_%d' % hs] = np.NaN
        new_pool['set_hash_%d' % hs] = np.NaN
        #new_pool['art_hash_%d' % hs] = np.NaN
    for ind, card_info in card_pool.iterrows():
        if ind % 100 == 0:
            print('Calculating hashes: %dth card' % ind)
@@ -60,28 +49,80 @@
        for card_name in card_names:
            # Fetch the image - name can be found based on the card's information
            card_info['name'] = card_name
            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' % (Config.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
            img_card = Image.fromarray(card_img)
            img_set = Image.fromarray(set_img)
            for hs in hash_size:
                card_hash = ih.phash(img_card, hash_size=hs)
                set_hash = ih.whash(img_set, hash_size=hs)
                card_info['card_hash_%d' % hs] = card_hash
                card_info['set_hash_%d' % hs] = 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

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 = 15
    num_partitions = round(card_pool.shape[0]/100)
    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)
@@ -197,7 +238,7 @@
    return corrected


def find_card(img, thresh_c=5, kernel_size=(3, 3), size_thresh=10000):
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
@@ -209,19 +250,29 @@
    # 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
    _, 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_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
@@ -238,7 +289,44 @@
        peri = cv2.arcLength(cnt, True)
        approx = cv2.approxPolyDP(cnt, 0.04 * peri, True)
        if size >= size_thresh and len(approx) == 4:
            cnts_rect.append(approx)
            # 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 %d' % i_cnt, 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]))
@@ -323,7 +411,7 @@
    img_result = img.copy()  # For displaying and saving
    det_cards = []
    # Detect contours of all cards in the image
    cnts = find_card(img_result, size_thresh=size_thresh)
    cnts = find_card(img_result, size_thresh=size_thresh, debug=debug)
    for i in range(len(cnts)):
        cnt = cnts[i]
        # For the region of the image covered by the contour, transform them into a rectangular image
@@ -342,15 +430,49 @@
        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')
        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' % hash_size]
            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))
        hash_diff = min_card['hash_diff']

        # Render the result, and display them if needed
        cv2.drawContours(img_result, [cnt], -1, (0, 255, 0), 2)
@@ -358,7 +480,7 @@
                    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 + ', ' + str(hash_diff), (0, 20),
            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:
@@ -371,7 +493,7 @@


def detect_video(capture, card_pool, hash_size=32, size_thresh=10000,
                 out_path=None, display=True, show_graph=True, debug=False):
                 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
@@ -388,8 +510,9 @@
    # 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)) + img_graph.shape[1]
        height = max(round(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)), img_graph.shape[0])
        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))
@@ -398,9 +521,15 @@
    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
@@ -408,7 +537,7 @@
                cv2.waitKey(0)
                break
            # Detect all cards from the current frame
            det_cards, img_result = detect_frame(frame, card_pool, hash_size=hash_size, size_thresh=size_thresh,
            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
@@ -430,18 +559,50 @@
                    else:
                        exist_cards[key] = exist_cards[key][1 - f_len:] + [0]
                    if len(val) == f_len and sum(val) == 0:
                        gone.append(key)
                        gone.append(key)  # not there anymore

                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

@@ -464,10 +625,17 @@
            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(args):
    # Specify paths for all necessary files

    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)
@@ -476,15 +644,23 @@
        # 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')
        calc_image_hashes(card_pool, save_to=pck_path)
        card_pool = calc_image_hashes(card_pool, save_to=pck_path, hash_size=hash_sizes)
    ch_key = 'card_hash_%d' % args.hash_size
    card_pool = card_pool[['name', 'set', 'collector_number', ch_key]]
    set_key = 'set_hash_%d' % args.hash_size
    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]]

    # 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
@@ -493,12 +669,16 @@
    # 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[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())

    # If the test file isn't given, use webcam to capture video
    if args.in_path is None:
        capture = cv2.VideoCapture(0)
        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)
        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)
                     display=args.display, show_graph=args.show_graph, debug=args.debug, crop_x=500, crop_y=200)
        capture.release()
    else:
        # Save the detection result if args.out_path is provided