The text detection methods based on deep learning has strong feature learning ability and generalization ability,but the inference speed is usually slow.To solve this problem,a fast horizontal text detection method T-YOLOv5 based on improved YOLOv5 is proposed.By em-bedding an improved CAM(channel attention module)in SPPF(spatial pyramid pooling-fast)module,the feature extraction ability of the network is improved,and the shape loss is added to the CIoU(complete IoU)loss to improve the convergence speed of the loss function.The F value of the proposed method reaches 86.5 on the public dataset ICDAR2013,and the inference speed reaches 112 FPS.Experimental results show that the proposed method T-YOLOv5 is competitive with the existing text detection methods based on deep learning in terms of detection results and inference speed.
维普
万方数据