Research On Multi-class Object Detection Method Based On Multi-spectral Remote Sensing Image

The research on object detection and classification methods in multi-spectral remote sensing images is a challenging and meaningful topic in the field of ground object detection.In recent years,deep neural network-based object detection and classification methods have made significant progress and achievements.At present,in response to the problem of multi class object detection and recognition in multi-spectral remote sensing images,this article intends to start with modules such as feature fusion,feature pyramid network,dynamic convolution neural network,and global attention mechanism to solve the technical problems faced by multi class object recognition technology in multi-spectral remote sensing images.(1)For complex environments and the characteristics of objects themselves,convolution neural networks cannot better learn the features of objects.Therefore,a faster-RCNNi object detection and classification method has been proposed.First,it mainly studies the spatial and Semantic information of learning objects,and enhances the shallow feature information to improve the recognition accuracy of small and dense objects.Secondly,in Faster-RCNN,Asymmetric Convolution Blocks are introduced into the network model to improve the expansion rate of the convolution kernel and its solving ability in the local domain.Finally,the correctness and applicability of the proposed method were demonstrated through experimental validation of the data set,as well as quantitative and qualitative analysis.(2)When existing deep learning object detection methods are used for object detection in multi-spectral remote sensing images,the detection results are not satisfactory.Exploring and improving object detection and classification methods for multi-spectral remote sensing images remains a challenging task.Therefore,a method for object detection and classification in multi-spectral remote sensing images based on feature pyramid networks was studied Proposed YOLOv3 i object detection and classification method.First,a new Feature Pyramid Network module is introduced to fuse the features of multiple Receptive field in the low-level feature map to solve the problem of insufficient effective information in the backbone feature extraction and weak information expression in the feature map.Secondly,a multi-level feature pyramid structure based on "top-down" is designed,and a smooth and non monotonic Mish activation function is introduced.Finally,based on experimental verification and quantitative and qualitative analysis of the multi-spectral remote sensing image datasets,the superiority of the proposed YOLOv3 i method was demonstrated.(3)Multi-spectral remote sensing images have complex backgrounds and small object scales.object detection methods based on natural images are difficult to achieve good results,especially for weak objects,which cannot achieve accurate positioning and recognition.Therefore,optimization research was conducted to improve feature extraction capabilities and object detection performance.A dynamic convolution neural network based object detection and classification method for multi-spectral remote sensing images was studied,and an Efficientnet-YOLOv3 i object detection and recognition method was proposed.Firstly,a Dynamic Convolution Neural Network module is introduced to address the issues of small effective information extracted by backbone feature extraction networks and weak information representation ability of feature maps.This module does not increase the depth and width of the network.Secondly,the performance of the model is improved by paying attention to aggregating multiple convolution kernels,where the convolution kernels share the same kernel size,input,and output dimensions.Finally,based on experimental verification and quantitative and qualitative analysis of the multi-spectral remote sensing image data set,the superiority of the proposed Efficientnet-YOLOv3 i method was demonstrated.(4)Because Multi-spectral Remote Sensing Images have the characteristics of dense target,uneven distribution,large scale change and complex background environment,it is very difficult to detect and classify targets..Therefore,in response to the problem of object detection and classification in multi-spectral remote sensing images,the Global Attention Mechanism and feature fusion are introduced into the object detection and classification methods of remote sensing images,and a YOLOXi object detection and classification method is proposed.First,aiming at the problem of small amount of effective information and weak information reconstruction of feature extraction in the backbone network The representation ability of feature maps introduces a Global Attention Mechanism module and uses a sequential channel spatial attention mechanism.The convolution block attention module aims to amplify the interactive features of the global dimension while reducing information dispersion.Secondly,a Feature Enhancement Module is constructed to fuse multiple Receptive field features in the low-level feature map to improve the object feature extraction capability of the backbone feature extraction network.Thirdly,under the action of Feature Fusion and global attention mechanism,FRe LU activation function is introduced.Finally,the superiority of the proposed YOLOXi method was demonstrated through experimental verification and quantitative and qualitative analysis of the multi-spectral remote sensing image data set.