def obj_inference_crop(obj_inferece, img, resize_scale):
roitransformer = obj_inferece
det = roitransformer.inference_single(img)
det = det[0]
# 如果某一类中出现了多个检测，那就取其中confidence最高的那一个检测，升序的[-1]，并且取前8个值为坐标
# print('det:{}'.format(det))
# assert det==np.empty((0,9))
if det.shape != (0, 9):
# print(det)
rgt_det = det[np.argmax(det[:, -1]), :-1]
# 根据长宽比进行调整
# xc, yc=np.mean(rgt_det[0::2]),np.mean(rgt_det[1::2])
_, reranked_nodes = re_rank(rgt_det) # re_rank之后是shape为(-1,2)的列表
reranked_nodes = np.array(reranked_nodes)
near_w = np.abs(reranked_nodes[0, 0] - reranked_nodes[1, 0])
near_h = np.abs(reranked_nodes[0, 1] - reranked_nodes[2, 1])
# 调整坐标和图片保证正立或倒立，但不倾倒
if near_h < near_w:
pass
else:
imgRotation, matRotation = DumpRotateImage(img, 90)
for i in range(len(reranked_nodes)):
points_arr = reranked_nodes[i, :]
reranked_nodes[i, :] = np.dot(
matRotation,
np.array([[points_arr[0]], [points_arr[1]],
[1]])).reshape(-1)
img = imgRotation
rgt_det = [reranked_nodes.reshape(-1)]
words = crop_words(img,
np.array(rgt_det) / resize_scale,
height=resize_scale,
grayscale=False)
return words[0]
else:
print('???')
# rgt_det = det
return None
def crop_boxes(self, result, grayscale=True):
img = self.resized_image
if len(result) > 0:
rboxes = result[:, 12:17]
boxes = np.asarray([rbox3_to_polygon(r) for r in rboxes])
rboxes = np.array([polygon_to_rbox(b) for b in np.reshape(boxes, (-1, 4, 2))])
bh = rboxes[:, 3]
# NOTE : without margin
# rboxes[:,2] += bh * 0.1
# rboxes[:,3] += bh * 0.2
boxes = np.array([rbox_to_polygon(f) for f in rboxes])
boxes = np.flip(boxes, axis=1)
boxes = np.reshape(boxes, (-1, 8))
# boxes_mask_a = np.array([b[2] > b[3] for b in rboxes]) # width > height, in square world
boxes_mask_b = 1 - np.array([(np.any(b < 0) or np.any(b > 1)) for b in boxes]) # box inside image
# added filter to filter out detections near edges
# NOTE : RoI is located on the centre of image
boxes_mask_c = 1 - np.array([(np.any(b[:2] < 0.2) or np.any(b[:2] > 0.8)) for b in boxes])
boxes_mask_c = np.array(boxes_mask_c)
# NOTE : fancier methods using numpy slicing didnt work
# had to do it old school way
# boxes_mask = np.array(boxes_mask_a* boxes_mask_b*boxes_mask_c)
boxes_mask = np.array(boxes_mask_b * boxes_mask_c)
boxes_ = []
for i, m in enumerate(boxes_mask):
if m == 1:
boxes_.append(boxes[i])
boxes = np.array(boxes_)
if len(boxes) == 0:
return -1, -1, -1
words = crop_words(img, boxes, self.input_height, width=self.input_width, grayscale=True)
words = np.asarray([w.transpose(1, 0, 2) for w in words])
return words, boxes, 1
else:
return -1, -1, -1
boxes = boxes[boxes_mask]
rboxes = rboxes[boxes_mask]
if len(boxes) == 0:
boxes = np.empty((0,8))
#for b in boxes:
# xy = b.reshape((-1,1,2)) / input_size * [vid_w, vid_h]
# xy = np.round(xy)
# xy = xy.astype(np.int32)
# cv2.polylines(img1, [xy], True, (0,0,255))
boxes = np.clip(boxes/512, 0, 1)
words = crop_words(img, boxes, input_height, width=input_width, grayscale=True)
words = np.asarray([w.transpose(1,0,2) for w in words])
if len(words) > 0:
res_crnn = rec_model.predict(words)
xy = rboxes[:,:2]
xy[:,0] = xy[:,0] - rboxes[:,2] / 2
xy = xy / input_size * [vid_w, vid_h]
for i in range(len(words)):
idxs = np.argmax(res_crnn[i], axis=1)
confs = res_crnn[i][range(len(idxs)),idxs]
non_blank_mask = idxs != len(alphabet)-1
if np.any(non_blank_mask):
boxes[:, 0::2] /= img_width
boxes[:, 1::2] /= img_height
boxes = np.concatenate([boxes, np.ones([boxes.shape[0], 1])], axis=1)
boxes = np.copy(boxes[:, :-1])
# drop boxes with vertices outside the image
mask = np.array([not (np.any(b < 0.) or np.any(b > 1.)) for b in boxes])
boxes = boxes[mask]
if len(boxes) == 0: continue
try:
words = crop_words(img, boxes, input_height, input_width, True)
except Exception as e:
import traceback
print(traceback.format_exc())
print(fname)
continue
mask = np.array([w.shape[1] > w.shape[0] for w in words])
words = [words[j] for j in range(len(words)) if mask[j]]
if len(words) == 0: continue
idxs_words = np.arange(len(words))
np.random.shuffle(idxs_words)
words = [words[j] for j in idxs_words]