Pruning Based Deep Network is Used for Text Detection of Natural Scene Images

The text in natural scene images contains a wealth of content, marking a large amount of information about this scene, which is of great value for image applications. Because of the blurring, distortion, warping and other problems of the text in natural scene, it is extremely difficult to detect it. The best tool for images processing is CNN, but the number of channel layers in recent CNN is getting deeper and deeper, most of them have hundreds of layers, which consumes a lot and requires high equipment. In this paper, the channel level is screened, and then the useless channel level is deleted. Combined with the latest Yolo v3 model, which has a good effect on small target detection, the text in natural scene is detected after pruning, and finally the performance is reflected on the ICDAR2015 data set. Experimental results show that the pruned model has the slightly lower precision than the original model, but the training time is greatly reduced.