Multi-class Object Detection Algorithm Based on Convolutional Neural Network

In order to improve the accurate recognition rate and localization rate of multi-class object detection, a new network structure, Res-YOLO-R., based on the combination of Residual Network (ResNet) and You Only Look Once (YOLO) detection network, is proposed. To improve the location ability and speed up the convergence of the network, the number and size of prediction boxes for YOLO network are redesigned by clustering analysis algorithm. Removing part of the pool layer and using convolution layer to raise or reduce the dimension of the feature to improve the ability of feature extraction and computing of the network. ResNet is designed as the feature extraction part, and the final average pool layer and the full connection layer are removed, and combines with the improved YOLO detection network to improve the degradation problem caused by the increasement of the network depth. In order to make the network learn object context information better, the ROUTE and REORG layers are used to fuse feature from different layers, and the feature map is reorganized. Through the comparison of experiments on commodity data sets, the network structure can effectively reduce the false detection rate and miss detection rate, improve the detection accuracy, positioning ability and recall rate of commodities, and have good real-time and generalization ability and strong practicability.