exam-segment-django/segment/sheet_resolve/analysis/exam_number/exam_number_row_column.py

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