Research On Model Optimization And Pruning Method For Object Detection In Remote Sensing Images

Using object detection technology to extract the category and location of the objects of interest from high-resolution remote sensing images has wide applications,including surveillance,resource exploration,etc.With the rapid development of deep learning,the accuracy of remote sensing image object detection improves greatly.To improve detection accuracy of typical objects in remote sensing images under complex scenes and meet real-time requirements,optimization of detection branch based on YOLOv3 model is done,and network pruning technology is used to achieve model compression in this paper.The main work of the paper is listed as follows:To solve the problems that the size of objects varies greatly and the number of small objects is large in remote sensing images,an improved YOLOv3 model for remote sensing image object detection based on weighting strategy is proposed.Firstly,the detection branch of feature map with smaller receptive fields is added to improve small-size object detection accuracy in remote sensing image.Secondly,an adaptive weighted fusion method for multi-scale feature maps is designed,which improves object detection accuracy by excavating representation capability of feature extraction network and fusing the multi-scale features comprehensively.Finally,a new remote sensing image object detection dataset containing four classes of objects in DIOR dataset is built,which is named as AVSS.The proposed model are trained and tested using AVSS dataset,and the model improves the m AP(mean Average Precision),which verifies the effectiveness of the proposed model.To further improve real-time performance of object detection on the basis of YOLOv3 model,a model compression method based on network pruning is proposed.Firstly,the parameters and computational cost of YOLOv3 model are reduced by reducing the number of feature channels of the convolutional layer in YOLOv3.This model is then pruned based on the scaling factor of the BN layer,which further reduces the complexity of the model.Using the AVSS dataset for experiment,the model after network pruning effectively reduces the parameter amount and model size,and significantly improves the real-time detection,which realizes the balance of detection accuracy,model capacity and detection speed.Finally,the influence of the input image size on the performance of the model is discussed,and experimental comparisons with other lightweight object detection models are carried out.Finally,the main work and further research of the paper is summarized.