A lightweight neural network model for security inspection targets

Although the application of the Yolo-V3 algorithm in target detection has a good performance, it is difficult to deploy the model to mobile platforms such as embedded devices due to the large network model and many redundant parameters. Aiming at the above limitations, this paper proposes a cyclic pruning algorithm based on O-T (Optimal Thresholding) threshold selection, prunes the original Yolo-V3 model, and obtains a lightweight model for security inspection target detection. Firstly, a local sparse strategy is proposed in the sparse training part to make the γ factor on which pruning depends more distinguishable. Secondly, in the threshold selection of the pruning process, the O-T threshold algorithm is introduced to make the selection of the pruning threshold more rational. Finally, in the fine-tuning part, an improved knowledge distillation strategy is proposed to improve the effect of precision recovery. The experimental results show that the pruning algorithm can achieve a higher pruning rate while ensuring a small loss of accuracy, and obtain a more lightweight security inspection target detection model.