Research And Implementation Of Remote Sensing Image Target Detection Algorithm Based On YOLOv

Remote sensing target detection aims to identify and locate targets in remote sensing images by using remote sensing and computer vision technology,which has important applications in the military,security,environment,resource management,and other fields.Although current target detection algorithms have made great progress,for remote sensing target detection tasks with complex backgrounds and little available texture information,they still face the following difficulties: firstly,the convolutional layer of the backbone network is too deep to lose remote sensing small target semantic information,resulting in remote sensing small target miss detection;secondly,the direct use of summation between different resolution feature layers for fusion leads to feature redundancy and under-utilization.Therefore,in order to improve the accuracy and speed of remote sensing target detection,this paper improves the backbone network feature extraction module and neck feature fusion module of the YOLOv4 model to make it more suitable for remote sensing target detection.The main research contents and innovative work are as follows:First of all,in order to solve the problem of losing the semantic information of remote sensing small targets in a backbone network with too many layers,this paper proposes to improve the YOLOv4 backbone network with a densely connected network(DenseNet)to enhance the feature extraction capability.The last four layers of each of the two eight-layer residuals of the model backbone network are replaced with the modified DenseNet,and the top ReLU activation function in the Denseblock is removed to ensure the subsequent network state;an expanded convolution is added to the Dense Block module to increase the perceptual field of the network;and a jump connection is used to deploy the network,which makes the neural network easier to train.easier to train.Experiments on the DOTA dataset show that the improved method improves the detection accuracy of remote sensing small targets.Secondly,to address the problem that the feature fusion module of the original YOLOv4 has insufficient feature information utilization and fusion capability,this paper proposes to use the improved weighted Bidirectional Feature PyramidNetwork(Bi FPN)as the feature fusion module to improve the feature fusion capability of the YOLOv4 network.The improved Bi FPN uses the resolution of each layer as the weighting factor,and the contribution of each layer feature to the network is specified during fusion to ensure that the network gradually learns the important information of each input feature during the training process.A Channel Attention Mechanism(CAM)was also added between the neck and head to improve sensitivity to image texture information,especially for targets with limited feature information.In addition,CDIoU and CDIoU loss were used instead to enhance the effect of bounding box regression.Experimental results show that the above improvements can improve the model’s ability to detect targets with limited feature information,and the number of detected targets is increased.Finally,this paper designs and implements a PyQt5-based software platform for remote sensing image target detection in order to demonstrate remote sensing target detection functions in a more user-friendly manner.