def model(img,
detectAngle=False,
config={},
leftAdjust=False,
rightAdjust=False,
alph=0.2,
ifadjustDegree=False):
"""
@@param:img,
@@param:ifadjustDegree 调整文字识别倾斜角度
@@param:detectAngle,是否检测文字朝向
"""
angle, degree, img = eval_angle(img,
detectAngle=detectAngle,
ifadjustDegree=ifadjustDegree)
if opencvFlag != 'keras':
img, f = letterbox_image(img, IMGSIZE)
else:
f = 1.0 ##解决box在原图坐标不一致问题
config['img'] = img
timeTake = time.time()
text_recs = text_detect(**config)
timeTake = time.time() - timeTake
print('Detect take:{}s'.format(timeTake))
newBox = sort_box(text_recs)
result = crnnRec(np.array(img), newBox, leftAdjust, rightAdjust, alph,
1.0 / f)
return img, result, angle
def test_keras_detect(img, **args):
scale = args.get('scale', 608)
maxScale = args.get('maxScale', 608)
boxes, scores = text_detect(img, scale, maxScale, prob=0.07)
boxes, scores = box_cluster(img, boxes, scores, **args)
boxes = sort_box(boxes)
leftAdjustAlph = args.get('leftAdjustAlph', 0.05)
rightAdjustAlph = args.get('rightAdjustAlph', 0.05)
tmp = plot_box(img, boxes)
tmp.save('./14_text_nms.jpg')
def model_PSENET(self, img, **args):
detectAngle = args.get('detectAngle', False)
if detectAngle:
img, angle = self.detect_angle(img)
else:
angle = 0
scale = args.get('scale', 608)
maxScale = args.get('maxScale', 608)
MAX_HORIZONTAL_GAP = args.get('MAX_HORIZONTAL_GAP', 15)
MIN_V_OVERLAPS = args.get('MIN_V_OVERLAPS', 0.01)
TEXT_PROPOSALS_MIN_SCORE = args.get('TEXT_PROPOSALS_MIN_SCORE', 0.9)
Adjustbox = args.get('Adjustbox', [-5, -5, 5, 5])
print('scale', scale, MAX_HORIZONTAL_GAP)
boxes = \
self.textModel(img, min_len=scale, max_len=maxScale, score_thre=TEXT_PROPOSALS_MIN_SCORE,
max_dist=MAX_HORIZONTAL_GAP, threshold_overlap_v=MIN_V_OVERLAPS, move_rect=Adjustbox)
import cv2
im = img.copy()
bo = []
for rt in boxes:
if not (rt[0] == rt[2] or rt[1] == rt[3]):
bo.append(rt)
cv2.rectangle(im, (rt[0], rt[1]), (rt[2], rt[3]), (0, 0, 255), 2)
boxes = np.array(bo)
im_show = cv2.resize(im, (im.shape[1] // 2, im.shape[0] // 2))
image = cv2.cvtColor(im_show, cv2.COLOR_BGR2RGB)
# cv2.imshow('img2', im_show)
# cv2.waitKey(0)
boxes_8 = np.zeros((len(boxes), 8), np.int32)
for i in range(len(boxes)):
boxes_8[i, 0] = boxes[i, 0] # x1
boxes_8[i, 1] = boxes[i, 1] # y1
boxes_8[i, 2] = boxes[i, 2] # x2
boxes_8[i, 3] = boxes[i, 1] # y1
boxes_8[i, 4] = boxes[i, 2] # x2
boxes_8[i, 5] = boxes[i, 3] # y2
boxes_8[i, 6] = boxes[i, 0] # x1
boxes_8[i, 7] = boxes[i, 3] # y2
# print(boxes_8,'boxes_8')
boxes = sort_box(boxes_8)
# print(boxes,'boxes')
leftAdjustAlph = args.get('leftAdjustAlph', 0)
rightAdjustAlph = args.get('rightAdjustAlph', 0)
res = self.ocr_batch(img, boxes, leftAdjustAlph, rightAdjustAlph)
return res, angle, image
def model(self, img, **args):
detectAngle = args.get('detectAngle', False)
if detectAngle:
img, angle = self.detect_angle(img)
else:
angle = 0
scale = args.get('scale', 608)
maxScale = args.get('maxScale', 608)
boxes, scores = self.detect_box(img, scale, maxScale) ##文字检测
boxes, scores = self.box_cluster(img, boxes, scores, **args)
boxes = sort_box(boxes)
leftAdjustAlph = args.get('leftAdjustAlph', 0)
rightAdjustAlph = args.get('rightAdjustAlph', 0)
res = self.ocr_batch(img, boxes, leftAdjustAlph, rightAdjustAlph)
return res, angle
def model(img, detectAngle=False, config={}, leftAdjust=False, rightAdjust=False, alph=0.2):
"""
@@param:img,
@@param:ifadjustDegree 调整文字识别倾斜角度
@@param:detectAngle,是否检测文字朝向
"""
angle, img = eval_angle(img, detectAngle=detectAngle) ##文字方向检测
if opencvFlag != 'keras':
img, f = letterbox_image(Image.fromarray(img), IMGSIZE) ## pad
img = np.array(img)
else:
f = 1.0 ##解决box在原图坐标不一致问题
config['img'] = img
text_recs = text_detect(**config) ##文字检测
newBox = sort_box(text_recs) ##行文本识别
result = crnnRec(np.array(img), newBox, leftAdjust, rightAdjust, alph, 1.0 / f)
return img, result, angle
def model(img,
file_name,
detectAngle=False,
config={},
leftAdjust=False,
rightAdjust=False,
alpha=0.2):
"""
@@param:img,
@@param:ifadjustDegree 调整文字识别倾斜角度
@@param:detectAngle,是否检测文字朝向
"""
angle, img = eval_angle(img, detectAngle=detectAngle) # 进行文字方向检测
if opencvFlag != 'keras':
img, f = letterbox_image(Image.fromarray(img), IMGSIZE) # pad
img = np.array(img)
else:
f = 1.0 # 解决box在原图坐标不一致问题
config['img'] = img
image_cv = copy.copy(img)
cv2.imshow("model", image_cv)
text_recs = text_detect(**config) # 文字检测
# print('text_recs', text_recs)
# draw_boxes(image_cv, text_recs, "text_recs", (255, 255, 0))
newBox = sort_box(text_recs) # 文本进行排序（从上到下进行排序）
print('newBox', newBox)
draw_boxes(image_cv, text_recs, file_name, (255, 0, 0), "./output")
# print('model', img.shape)
result = crnnRec(np.array(img), newBox, leftAdjust, rightAdjust, alpha,
1.0 / f) # ocr识别
print('result', result)
return img, result, angle
def model(img,
detectAngle=False,
config={},
leftAdjust=False,
rightAdjust=False,
alph=0.2,
bili=1.2):
"""
@@param:img,
@@param:ifadjustDegree 调整文字识别倾斜角度
@@param:detectAngle,是否检测文字朝向
"""
print(22222222222)
'''
这个地方需要加入图片预处理!!!!!!滤波等.
确实这个算法没有加入,需要加入!!!!!!!!!!!
'''
angle, img = eval_angle(img, detectAngle=detectAngle) ##文字方向检测
if opencvFlag != 'keras':
img, f = letterbox_image(Image.fromarray(img), IMGSIZE) ## pad
img = np.array(img)
else:
f = 1.0 ##解决box在原图坐标不一致问题
config['img'] = img
config['bili'] = bili
text_recs, scores, boxForSingleAfterNMS, scoresForSingle, keepIndForSingle, tp_groups, Allboxes, Allscores = text_detect(
**config) ##文字检测
newBox, tp_groups = sort_box(text_recs,
tp_groups) #按照列高排序,符合我们阅读顺序! ##下行行文本识别
print(newBox)
result = crnnRec(np.array(img), newBox, leftAdjust, rightAdjust, alph,
1.0 / f, tp_groups, boxForSingleAfterNMS, scoresForSingle)
return img, result, angle, scores, text_recs, newBox, boxForSingleAfterNMS, scoresForSingle, keepIndForSingle, tp_groups, Allboxes, Allscores
def net_output_process(batch_preds,
batch_shape,
batch_shape_padded,
prob=0.05):
"""
将主干网络的批输出转换为boxes,scores,这里方便兼容之前代码，
暂且使用for循环,以后可替换为vectorize处理
@params batch_preds(list of arrays):list长度代表n个采样尺度，其中每个\
array形状为(batch_size,grid_size_w,grid_size_h,3*(4+1+num_classes))
@params batch_shape(list of tuples):图片原始长宽
@params prob(float):置信度小于prob的box将被忽略
@returns batch_boxes(array): [字符区域数量，8]
@returns batch_scores:[字符区域数量,]
"""
batch_boxes = []
batch_scores = []
MAX_HORIZONTAL_GAP = 100
MIN_V_OVERLAPS = 0.6
MIN_SIZE_SIM = 0.6
textdetector = TextDetector(MAX_HORIZONTAL_GAP, MIN_V_OVERLAPS,
MIN_SIZE_SIM)
TEXT_PROPOSALS_MIN_SCORE = 0.1
TEXT_PROPOSALS_NMS_THRESH = 0.3
TEXT_LINE_NMS_THRESH = 0.99
LINE_MIN_SCORE = 0.1
leftAdjustAlph = 0.01
rightAdjustAlph = 0.01
# 首先初步对模型主干输出进行预处理
for y1, y2, y3, image_shape, input_shape in zip(batch_preds[0],
batch_preds[1],
batch_preds[2],
batch_shape,
batch_shape_padded):
outputs = [y1, y2, y3, image_shape, input_shape]
box, scores = box_layer(outputs, anchors, num_classes)
h, w = image_shape
keep = np.where(scores > prob)
# box[:, 0:4][box[:, 0:4]<0] = 0
box = np.array(box)
scores = np.array(scores)
box[box < 0] = 0
box[:, 0][box[:, 0] >= w] = w - 1
box[:, 1][box[:, 1] >= h] = h - 1
box[:, 2][box[:, 2] >= w] = w - 1
box[:, 3][box[:, 3] >= h] = h - 1
boxes = box[keep[0]]
scores = scores[keep[0]]
# 筛选出需要的box，并且进行nms，字符行组合
boxes, scores = textdetector.detect(boxes, scores[:, np.newaxis],
(h, w), TEXT_PROPOSALS_MIN_SCORE,
TEXT_PROPOSALS_NMS_THRESH,
TEXT_LINE_NMS_THRESH,
LINE_MIN_SCORE)
boxes = sort_box(boxes)
batch_boxes.append(boxes)
batch_scores.append(scores)
# print('done')
return batch_boxes, batch_scores
def model_CRAFT(self, img, **args):
detectAngle = args.get('detectAngle', False)
if detectAngle:
img, angle = self.detect_angle(img)
else:
angle = 0
scale = args['scale']
MAX_HORIZONTAL_GAP = args['MAX_HORIZONTAL_GAP']
MIN_V_OVERLAPS = args['MIN_V_OVERLAPS']
TEXT_PROPOSALS_MIN_SCORE = args['TEXT_PROPOSALS_MIN_SCORE']
Adjustbox = args['Adjustbox']
pixel_filter = args['pixel_filter']
batch_by_1 = args['batch_by_1']
scoremap_enhance_pixel = args['scoremap_enhance_pixel']
boxes = \
self.textModel(img, image_inference_scale=scale, score_thre=TEXT_PROPOSALS_MIN_SCORE, batch_by_1=batch_by_1,
max_dist=MAX_HORIZONTAL_GAP, threshold_overlap_v=MIN_V_OVERLAPS, move_rect=Adjustbox,
pixel_filter=pixel_filter,scoremap_enhance_pixel=scoremap_enhance_pixel)
bo = []
for rt in boxes:
Xs = [rt[0], rt[2], rt[4], rt[6]]
Ys = [rt[1], rt[3], rt[5], rt[7]]
x1 = min(Xs)
x2 = max(Xs)
y1 = min(Ys)
y2 = max(Ys)
hight = y2 - y1
width = x2 - x1
# print(hight, '-------------------', width)
if hight <= 10 or hight >= 300:
pass
else:
bo.append(rt)
boxes = sort_box(bo)
leftAdjustAlph = args.get('leftAdjustAlph', 0)
rightAdjustAlph = args.get('rightAdjustAlph', 0)
# for rt in boxes:
# cv.rectangle(img, (rt[0], rt[1]), (rt[2], rt[3]), (0, 0, 255), 2)
res = self.ocr_batch(img, boxes, leftAdjustAlph, rightAdjustAlph)
import cv2
im = img.copy()
for pts in boxes:
# print(pts)
pts = pts.reshape((-1, 1, 2))
# print()
# print(pts)
cv2.polylines(im, [pts], True, (0, 0, 255), 1)
# cv2.imshow('img',im)
# cv2.waitKey()
# cv.imwrite('label.jpg',im)
return res, angle, im
def crnnRec(im, boxes, leftAdjust=False, rightAdjust=False, alph=0.2, f=1.0):
"""
crnn模型，ocr识别
@@model,
@@converter,
@@im:Array
@@text_recs:text box
@@ifIm:是否输出box对应的img
"""
results = []
# print("orcModel:",ocrModel)
# print("ocrFlag:",ocrFlag)
im = Image.fromarray(im)
# print("boxes:",boxes)
boxes = sort_box(boxes) # 没有排序
# print("boxes:",boxes)
import time
i = 1
for index, box in enumerate(boxes):
start_time = time.time()
degree, w, h, cx, cy = box
# partImg, newW, newH = rotate_cut_img(im, 90 + degree , cx, cy, w, h, leftAdjust, rightAdjust, alph)
partImg = crop_rect(im, ((cx, cy), (h, w), degree))
newW, newH = partImg.size
# partImg.thumbnail(newW*2,newH*2)
# partImg_array = np.uint8(partImg)
# if newH > 1.5 * newW:
# partImg_array = np.rot90(partImg_array, 1)
# partImg = Image.fromarray(partImg_array).convert("RGB")
# partImg.save("./debug_im/{}.jpg".format(index))
# angel_index = angle_handle.predict(partImg_array)
#
# angel_class = lable_map_dict[angel_index]
# # print(angel_class)
# rotate_angle = rotae_map_dict[angel_class]
#
# if rotate_angle != 0:
# partImg_array = np.rot90(partImg_array, rotate_angle // 90)
# partImg, box = rotate_cut_img(im, box, leftAdjust, rightAdjust)
# partImg = Image.fromarray(partImg_array).convert("RGB")
# partImg.save("./debug_im/{}.jpg".format(index))
# partImg.save(r'outputs/vis_invoice/{}.png'.format(index))
partImg_ = partImg.convert('L')
try:
# if crnn_vertical_handle is not None and angel_class in ["shudao", "shuzhen"]:
#
# simPred = crnn_verticv2.cvtColor(cal_handle.predict(partImg_)
# else:
pre = time.time()
simPred = crnn.predict(partImg_)
# print("simPred:",simPred)##识别的文本
# simPred = recognizer(partImg_)
except:
continue
if simPred[0].strip() != []:
# results.append({'cx': box['cx'] * f, 'cy': box['cy'] * f, 'text': simPred, 'w': box['w'] * f, 'h': box['h'] * f,
# 'degree': box['degree']})
# f 默认为 1
results.append({
'cx': cx * f,
'cy': cy * f,
'text': simPred[0],
'candidate': simPred[1],
'w': newW * f,
'h': newH * f,
'degree': degree
})
# print("results:",results)
i += 1
return results