import numpy as np import tensorflow as tf from segment.sheet_resolve.lib.ssd_model.utils import label_map_util, ops as utils_ops from segment.sheet_resolve.tools import tf_settings from segment.sheet_resolve.tools.tf_sess import SsdSess from PIL import Image import math tf_sess_dict = { 'exam_number_ssd': SsdSess('exam_number_ssd'), } exam_number_sess = tf_sess_dict['exam_number_ssd'] sess = exam_number_sess.sess detection_graph = exam_number_sess.graph def load_image_into_numpy_array(image): # print(image) image = image.convert('RGB') (im_width, im_height) = image.size return np.array(image.getdata()).reshape((im_height, im_width, 3)).astype(np.uint8) def run_inference_for_single_image(image): ops = detection_graph.get_operations() all_tensor_names = {output.name for op in ops for output in op.outputs} tensor_dict = {} for key in [ 'num_detections', 'detection_boxes', 'detection_scores', 'detection_classes', 'detection_masks' ]: tensor_name = key + ':0' if tensor_name in all_tensor_names: tensor_dict[key] = detection_graph.get_tensor_by_name( tensor_name) if 'detection_masks' in tensor_dict: # The following processing is only for single image detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0]) detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0]) # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size. real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32) detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1]) detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1]) detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks( detection_masks, detection_boxes, image.shape[0], image.shape[1]) detection_masks_reframed = tf.cast( tf.greater(detection_masks_reframed, 0.5), tf.uint8) # Follow the convention by adding back the batch dimension tensor_dict['detection_masks'] = tf.expand_dims( detection_masks_reframed, 0) image_tensor = detection_graph.get_tensor_by_name('image_tensor:0') # Run inference # start = time.time() output_dict = sess.run(tensor_dict, feed_dict={image_tensor: np.expand_dims(image, 0)}) # print(time.time()-start) # all outputs are float32 numpy arrays, so convert types as appropriate output_dict['num_detections'] = int(output_dict['num_detections'][0]) output_dict['detection_classes'] = output_dict[ 'detection_classes'][0].astype(np.uint8) output_dict['detection_boxes'] = output_dict['detection_boxes'][0] output_dict['detection_scores'] = output_dict['detection_scores'][0] if 'detection_masks' in output_dict: output_dict['detection_masks'] = output_dict['detection_masks'][0] return output_dict def image_detect(image_np, category, score_threshold): image_np = load_image_into_numpy_array(image_np) detections = [] w, h = image_np.shape[1], image_np.shape[0] output_dict = run_inference_for_single_image(image_np) boxes = output_dict['detection_boxes'] scores = output_dict['detection_scores'] labels = output_dict['detection_classes'] indices = np.where(scores > score_threshold) image_scores = scores[indices] image_boxes = boxes[indices] image_labels = labels[indices] image_detections = np.concatenate( [image_boxes, np.expand_dims(image_scores, axis=1), np.expand_dims(image_labels, axis=1)], axis=1) for detection in image_detections: y0 = int(detection[0] * h) x0 = int(detection[1] * w) y1 = int(detection[2] * h) x1 = int(detection[3] * w) label_index = int(detection[5]) label_name = category[label_index]['name'] detections.append((x0, y0, x1, y1, label_index, detection[4], label_name)) return detections def get_exam_number_row_and_col(left, top, image): im_resize = 512 ''' exam_number resize to 512*512''' image_src = Image.fromarray(image) if image_src.mode == 'RGB': image_src = image_src.convert("L") w, h = image_src.size if h > w: image_src = image_src.resize((int(im_resize / h * w), im_resize)) else: image_src = image_src.resize((im_resize, int(im_resize / w * h))) w_, h_ = image_src.size image_512 = Image.new(image_src.mode, (im_resize, im_resize), (255)) image_512.paste(image_src, [0, 0, w_, h_]) n_z = "0123456789" category_index = label_map_util.create_category_index_from_labelmap(tf_settings.exam_number_ssd_label, use_display_name=True) detections = image_detect(image_512, category_index, 0.5) if len(detections): box_xmin = [] box_ymin = [] box_xmax = [] box_ymax = [] x_distance_all = [] y_distance_all = [] x_width_all = [] y_height_all = [] all_small_coordinate = [] border = {} exam_number_ssd = {} ssd_column = 1 ssd_row = 1 for index, box in enumerate(detections): box0 = round(box[0] * (w / w_)) # Map to the original image box1 = round(box[1] * (h / h_)) box2 = round(box[2] * (w / w_)) box3 = round(box[3] * (h / h_)) if box[-1] == 'border': border = {'xmin': box0, 'ymin': box1, 'xmax': box2, 'ymax': box3 } # if box[2] - box[0] > 80 or box[3] - box[1] >80: # continue else: box_xmin.append(box0) box_ymin.append(box1) box_xmax.append(box2) box_ymax.append(box3) small_coordinate = {'xmin': box0 + left, 'ymin': box1 + top, 'xmax': box2 + left, 'ymax': box3 + top} all_small_coordinate.append(small_coordinate) x_width = box2 - box0 y_height = box3 - box1 x_width_all.append(x_width) y_height_all.append(y_height) sorted_xmin = sorted(box_xmin) sorted_ymin = sorted(box_ymin) sorted_xmax = sorted(box_xmax) sorted_ymax = sorted(box_ymax) # print(sorted_xmin, sorted_ymin) x_width_all_sorted = sorted(x_width_all, reverse=True) y_height_all_sorted = sorted(y_height_all, reverse=True) len_x = len(x_width_all) len_y = len(y_height_all) x_width_median = np.median(x_width_all_sorted) y_height_median = np.median(y_height_all_sorted) for i in range(len(sorted_xmin) - 1): x_distance = sorted_xmin[i + 1] - sorted_xmin[i] y_distance = sorted_ymin[i + 1] - sorted_ymin[i] if x_distance > (x_width_median - 5): ssd_column = ssd_column + 1 x_distance_all.append(x_distance) if y_distance > (y_height_median - 5): ssd_row = ssd_row + 1 y_distance_all.append(y_distance) # del the borders where small items are too large if x_width_all_sorted[i] - x_width_median > x_width_median: ssd_column = ssd_column - 1 elif x_width_median - x_width_all_sorted[i] > x_width_median: ssd_column = ssd_column - 1 if y_height_all_sorted[i] - y_height_median > y_height_median: ssd_row = ssd_row - 1 elif y_height_median - y_height_all_sorted[i] > y_height_median: ssd_row = ssd_row - 1 # Add rows and columns that might be missed x_distance_all_sorted = sorted(x_distance_all, reverse=True) y_distance_all_sorted = sorted(y_height_all, reverse=True) len_x_distance = len(x_distance_all) len_y_distance = len(y_distance_all) x_distance_median = np.median(x_distance_all_sorted) y_distance_median = np.median(y_distance_all_sorted) for i in range(len_x_distance): if x_distance_all[i] > 2 * x_distance_median - 4: ssd_column = ssd_column + 1 for i in range(len_y_distance): if y_distance_all[i] > 2 * y_distance_median - 4: ssd_row = ssd_row + 1 if ssd_row < 10: test = math.ceil(len_y / ssd_column) if test > ssd_row: ssd_row = test if ssd_row > 10: ssd_row = 10 average_height = int(np.mean(y_height_all)) average_width = int(np.mean(x_width_all)) location_ssd = {'xmin': sorted_xmin[0] + left, 'ymin': sorted_ymin[0] + top, 'xmax': sorted_xmax[-1] + left, 'ymax': sorted_ymax[-1] + top} exam_number_ssd = {'bounding_box': location_ssd, "single_height": average_height, "single_width": average_width, "rows": ssd_row, "cols": ssd_column, "option": n_z[:ssd_row].replace('', ',')[1:-1], "direction": 180, 'class_name': 'exam_number_col_row', 'all_small_coordinate': all_small_coordinate } else: exam_number_ssd = {} return exam_number_ssd