def table_detect(img, sc=(416, 416), thresh=0.5, NMSthresh=0.3):
"""
表格检测
img:GBR
"""
scale = sc[0]
img_height, img_width = img.shape[:2]
inputBlob, fx, fy = letterbox_image(img[..., ::-1], (scale, scale))
inputBlob = cv2.dnn.blobFromImage(inputBlob,
scalefactor=1.0,
size=(scale, scale),
swapRB=True,
crop=False)
tableDetectNet.setInput(inputBlob / 255.0)
outputName = tableDetectNet.getUnconnectedOutLayersNames()
outputs = tableDetectNet.forward(outputName)
class_ids = []
confidences = []
boxes = []
for output in outputs:
for detection in output:
scores = detection[5:]
class_id = np.argmax(scores)
confidence = scores[class_id]
if confidence > thresh:
center_x = int(detection[0] * scale / fx)
center_y = int(detection[1] * scale / fy)
width = int(detection[2] * scale / fx)
height = int(detection[3] * scale / fy)
left = int(center_x - width / 2)
top = int(center_y - height / 2)
if class_id == 1:
class_ids.append(class_id)
confidences.append(float(confidence))
xmin, ymin, xmax, ymax = left, top, left + width, top + height
xmin = max(xmin, 1)
ymin = max(ymin, 1)
xmax = min(xmax, img_width - 1)
ymax = min(ymax, img_height - 1)
boxes.append([xmin, ymin, xmax, ymax])
boxes = np.array(boxes)
confidences = np.array(confidences)
if len(boxes) > 0:
boxes, confidences = nms_box(boxes,
confidences,
score_threshold=thresh,
nms_threshold=NMSthresh)
boxes, adBoxes = fix_table_box_for_table_line(boxes, confidences, img)
return boxes, adBoxes, confidences
def table_detect(img, sc=(416, 416), thresh=0.5, NMSthresh=0.3):
"""
表格检测
:param img: GBR, 要检测的图片
:param sc: 预处理后图像的目标尺寸，一般有几个建议的值
:param thresh: 置信度阈值，大于此置信度的才保留
:param NMSthresh: 极大值抑制阈值
:return:
"""
scale = sc[0]
#获取img的前2位，图片的高度和宽度
img_height, img_width = img.shape[:2]
# 输入的Blob bbox, 新的宽度和原宽度的比值, 新的高度和原高度的比值
inputBlob, fx, fy = letterbox_image(img[..., ::-1], (scale, scale))
# 对输入图像进行预处理，均值，缩放，通道交互[H,W,C]-->[B,C,H,W]
inputBlob = cv2.dnn.blobFromImage(inputBlob, scalefactor=1.0, size=(scale, scale), swapRB=True, crop=False);
#设置模型的输入
tableDetectNet.setInput(inputBlob / 255.0)
# 返回没有连接的输出的layer的名字，
outputName = tableDetectNet.getUnconnectedOutLayersNames()
# 运行前向计算，计算OutputName的layers的输出, outputs输出结果的列表
outputs = tableDetectNet.forward(outputName)
#存放类别id，置信度，bbox
class_ids = []
confidences = []
boxes = []
#对于多个输出结果过滤
for output in outputs:
#处理每个结果, detection输出格式是[centerx,centery,w,h,xxxx, class1_confidence, class2_confidence]
# centerx 是bbox中心点坐标，w，h是bbox的宽和高
for detection in output:
#第5个和第6个是对每个类别的预测的置信度
scores = detection[5:]
#置信度最大的index是对应的是类别id
class_id = np.argmax(scores)
#获取对应的置信度
confidence = scores[class_id]
#检查置信度是否大于阈值
if confidence > thresh:
#还原到原图像的x，y，w，h
center_x = int(detection[0] * scale / fx)
center_y = int(detection[1] * scale / fy)
width = int(detection[2] * scale / fx)
height = int(detection[3] * scale / fy)
#bbox左顶点（x，y），这里用left是x，top是y
left = int(center_x - width / 2)
top = int(center_y - height / 2)
# 如果类别id是1
if class_id == 1:
class_ids.append(class_id)
confidences.append(float(confidence))
#计算bbox左上角和右下角的点的坐标
xmin, ymin, xmax, ymax = left, top, left + width, top + height
xmin = max(xmin, 1)
ymin = max(ymin, 1)
xmax = min(xmax, img_width - 1)
ymax = min(ymax, img_height - 1)
boxes.append([xmin, ymin, xmax, ymax])
#bboxes的列表
boxes = np.array(boxes)
#对应的confidences列表
confidences = np.array(confidences)
#NMS非极大值抑制过滤bbox
if len(boxes) > 0:
boxes, confidences = nms_box(boxes, confidences, score_threshold=thresh, nms_threshold=NMSthresh)
boxes, adBoxes = fix_table_box_for_table_line(boxes, confidences, img)
return boxes, adBoxes, confidences