Improved Yolo v3 via feature map design and anchor box design

Abstract

We introduced a method to improve the feature map and Anchor box of Yolo V3 network on VOC data set, so as to improve the detection accuracy of target in specific data set. Meanwhile, we also improved the detection accuracy of small targets to some extent. We improved Yolo V3 network. We increased the number of feature extraction layers and deepened its network structure, so that it could better detect targets. For different data sets, we needed to adjust the size of Anchor box. We added the extraction method of some feature layers of SSD into Yolo V3 network to match relative data sets. At the same time, we use the characteristics of ResNet to solve the problem of small target distortion after multiple convolution. The modified model improved mAP by 2.93% on VOC2007.We put forward three points (1) To change the original network structure, increase the number of feature layers and add convolution layers to deepen the network depth. (2) Change the scale of the prior box to match the target scale of the data set. (3) Deepen ResNet in shallow network to extract small targets, select the appropriate size of Anchor box according to specific small targets, enhance small target data and then conduct training.