YOLOV3 FUSION PRUNING BASED ON REMOTE SENSING IMAGE

In order to enable the YOLOv3 algorithm to perform real-time target detection on remote sensing equipment, model compression is a common solution. According to the characteristics of the loss of accuracy after the model is compressed, this paper added the attention mechanism algorithm to the model and trained the model. On this basis, we proposed a channel pruning method based on the fusion of the convolutional layer and the BatchNormal layer. Channel pruning was performed on YOLOv3, and we obtained the compressed YOLOv3 target detection model. After fine-tuning the pruned model, the accuracy of the model was restored. The experimental results show that the method of pruning after fusion in this paper reduces the size of the YOLOv3 model by 94.93% and increases the detection speed by 150.6% with only 0.6% mAP loss. The experiment proves that the model can be applied to real-time performance remote sensing image target detection with higher and higher detection accuracy, and suitable for remote sensing equipment with small storage space.