YOLOv3 Remote Sensing Image Object Detection With Auxiliary Networks

Remote sensing image target detection has been widely used in military and livelihood fields in recent years,but it is limited by the fact that most remote sensing targets are relatively small and do not have enough features for the existing deep learning networks to learn,so it is especially important to enhance the feature extraction capability of the existing deep learning networks.The innovative work is as follows:First,a dual-channel feature extraction network based on YOLOv3 is proposed,which aims to be able to enhance the feature extraction capability of YOLOv3.The auxiliary network is connected in parallel with the YOLOv3 backbone network,but the convolution kernel of the auxiliary network is different from that of the backbone network,which facilitates the extraction of diverse features to increase the variety of feature outputs.In order to make the features of the auxiliary network easier to integrate,a compression-incentive attention mechanism is used in the tail of the auxiliary network,which allows the network to learn the channel features of a specific target and serves to integrate the features of the auxiliary network.Secondly,an improved YOLOv3 network with an auxiliary network is proposed,and a Convolutional Block Attention Mechanism with better results is used in order for the network to learn spatial and channel features.The output of YOLOv3 is hierarchical,in other words,YOLOv3 cannot fuse the deep semantic information with the shallow detail information,which greatly reduces the ability of feature expression,so the Adaptively Spatial Feature Fusion method is used in this paper to make the features more fully fused.The loss function of YOLOv3 has a slow convergence speed and a long training time,so the DIo U loss function with a faster and more stable convergence speed is used in this paper,and a comparison test is made for all the above improvements,which proves that the improvements are effective.Finally,a PyQt5-based software platform for remote sensing image target detection is designed and implemented to make the method in this paper more convenient and intuitive to be reproduced.