The Improvement Research Of Multi-Target Detection Algorithm In Remote Sensing Images Of Complex Backgrounds

The detection of large traffic targets in remote sensing images is widely used,but due to the complexity and multi-angle nature of small targets in remote sensing images,the detection of small targets in remote sensing images is generally poor and technically difficult.Aiming at multi-target detection in complex environment of remote sensing images,a target detection algorithm based on SSD(Single Shot Multi BoX Detector)algorithm fused with multi-layer separation convolution module is proposed to improve the detection accuracy of multi-scale and dense target remote sensing large traffic targets.SSD network is used to detect a variety of large traffic targets,and for the shortcomings of SSD network for small target detection accuracy,the backbone network of SSD algorithm is replaced with HS-ResNet50(Hierarchical-Split Block on Convolutional Neural Network)for detailed feature extraction,and the backbone feature extraction network is further optimized in combination with multi-layer separation module to enhance the complex background dense target feature extraction capability.For the problem of loss function optimization defect of SSD network,GIoU(Generalized Intersection over Union)is used to locate the final prediction frame.The main research topics of this study are as follows:(1)In order to improve the problem that the anchor frame of the original SSD algorithm does not match the data set selected for study in this paper,the K-means clustering algorithm is to be used to generate a new anchor frame on the pre-selected data set in this paper,to optimize the original border loss function,and to replace the original IoU(Intersection over Union)with the generalized intersection and ratio GIoU to achieve an accurate reflection of the overlap between the anchor frame and the border frame and to enhance the accuracy of the target detection results.(2)The publicly available remote sensing image dataset contains a high diversity of targets,and not all of them are large vehicles,so it is necessary to pre-process the publicly available remote sensing image dataset to finally obtain the dataset used in the experiments of the algorithm in this paper.This effectively improves the experimental accuracy and the stability of the algorithm model.(3)The original SSD algorithm is prone to the problem of missing and false detection for detecting small targets.The backbone network in the original SSD algorithm is VGG16(Visual Geometry Group Network),which convolves through iterative algorithms,and although the speed is high enough,it is prone to the problem of gradient explosion and gradient disappearance.In this paper,we construct the backbone network HS-Resnet50 to replace the backbone network VGG16 in the original SDD algorithm by studying HS-Resnet and Resnet50 residual network to improve the detection accuracy of the overall algorithm while ensuring the speed.By conducting experimental comparison and analysis,the improved algorithm in this paper is experimented on UCAS-AOD and RSOD public remote sensing datasets,and the final results verify that compared with the original SSD algorithm,the proposed algorithm in this paper improves the mAP(mean Average Precision)values by 9.84%and 7.32%when the IoU threshold is 0.75 and 0.5,and achieves the optimized small target detection effect of the network.The experimental results show that the improved algorithm based on the backbone network optimization and anchor frame selection optimization results of the SSD target detection algorithm proposed in this paper can accomplish the task of detecting a variety of small targets in remote sensing images with good accuracy and speed.