Abstract: Real time object detection is a key technology in remote sensing field. The detection rate of object detection models based on deep network is high, but this kind of algorithm model often has lots of parameters and computational load, which makes it difficult to deploy the model on edge devices. Lightweight of this kind of model is a problem for the current deep network object detection algorithm. Based on the YOLOv5s object detection model, a convolution kernel pruning method is applied to the YOLOv5s model, first thinning the BN layer scaling factor, then pruning and fine tuning the convolution kernel, and training and testing with remote sensing aircraft dataset. The experimental results show that the object detection performance of the model can change within 2% when 30-50% of the model parameters are clipped. The method can effectively reduce the over fitting of YOLOv5s model and achieve the effect of reducing the size of the model.

Key words: remote sensing image, target detection, YOLOv5, pruning