Lightweight optimization of Convolutional Neural Network for multiobjective recognition in structured scene

There are many redundant parameters in Convolutional Neural Network(CNN) when it is used for target recognition in a specific scene, which will greatly occupy the calculation amount and affect the operating efficiency of software and hardware, and cannot meet the real-time target detection requirements of the algorithm in a specific scene. In this paper, channel pruning, layer pruning and their hybrid pruning experiments were carried out on you only look once version 3 (YOLOv3), a typical CNN target recognition model. The result of the hybrid pruning can greatly reduce the model parameters and the amount of calculation, and can reduce the model of resource utilization through the comparative analysis. And the volume of the model after hybrid pruning was reduced 94.4% when mean Average Precision (mAP) only loss 0.9%. The model inference time was reduced by 36.6%. This study could provide references for the optimization of object recognition model in structured scenes such as road and workshop.