Compression Model for Remote Sensing Image Target Detection Based on YOLOv3-SPP

‍ ‍In recent years， convolutional neural network （CNN） models have been widely applied in remote sensing image object detection tasks. However， the target characteristics of natural scene images and remote sensing images differ， and models designed for natural scenes often perform poorly in remote sensing image tasks. Additionally， many remote sensing image processing tasks need to be carried out on resource-limited platforms such as satellites and aircraft， making it difficult to deploy complex models with large parameter and computational requirements. To address these issues， this paper proposes adaptive improvements and parameter compression for the YOLOv3-SPP model， which has shown outstanding performance in natural scenes. First， an improved weighted pruning algorithm based on L1 norm and mean difference is proposed to better preserve important channels. Next， the pruned subnet collection is quickly evaluated， and the best-performing subnet is fine-tuned. In the pre-training and fine-tuning stages， the maximum pooling layer in the SPP module is replaced with a softmax weighted pooling layer to highlight features with larger weights in the deep network and improve the detection accuracy of the model.Experiments conducted on multiple public remote sensing datasets show that the improved YOLOv3-SPP model has better performance in remote sensing object detection tasks. Additionally， the pruning algorithm in this paper can reduce the model’s performance loss under the same parameter compression ratio.