5 / 2025-03-26 20:04:08

An Improved Pruning Method with Sparsification of Foreign Object Detection Model for Belt Conveyor

visual analysis,BN layer weight,pruning algorithm,model deployment

The belt conveyors with foreign object detection system need to satisfy terminal reasoning standards. In order to realize terminal reason of the conveyor belt foreign object detection model, an improved dynamic sparse pruning method based on network visualization analysis is proposed. This work introduces Gradient Class Activation Map (Grad-CAM), a visual analysis tool for the conveyor belt foreign object detection model, which can be used to identify the network structure that needs to be pruned. Additionally, a dynamic sparse pruning algorithm was presented, which relies on the Batch Normalization (BN) layers weights. The comparison experiments of the foreign object detection model and the foreign object terminal reasoning test experiment under embedded equipment are carried out. According to the experimental results, the pruning scale factor is set to 0.3 or 0.4, respectively. The number of model parameters is lowered by 26.99% and 41.63%. When deploying the pruned model to the Jetson Xavier NX and accelerating it by TensorRT, the model degrades the accuracy by 0.31% on average.