Research On Object Classification And Detection Method Based On Network Enhanced Model And Attention Fusion Mechanism

With the maturity of the high-resolution satellite series layout,people can more conveniently obtain rich and diverse high-resolution remote sensing images(hereinafter referred to as high-resolution images),which accelerates the application of remote sensing technology to agriculture,military,transportation,and ecological environment.The classification and detection of high-resolution images have always been a research hotspot in the application of remote sensing technology.Due to the wide variety of objects in highresolution images and different scales,the same object has different sizes,and different object have similar shapes,feature extractors are often designed manually for specific tasks,only shallow feature information such as color,texture,gradient,etc.can be extracted,which cannot effectively improve the accuracy of image analysis.How to improve the feature extraction method of high-resolution images and make full use of high-resolution images feature information is the key to improve the accuracy of image classification and object detection.Deep learning technology has been widely used in high-score image classification and object detection related fields.Its obvious advantage is that it can automatically extract highlevel semantic information of images.However,the current deep learning methods pay too much attention to the application of the deep feature information extracted by the network,ignore the basic information such as the position and texture of the shallow feature layer,at the same time,ignore data-driven characteristics of deep learning,samples can determine whether the model has the ability to be more robust and generalize.Therefore,based on the current deep learning framework,this paper improves the effect of high-score image classification and object detection by designing clever feature extraction modules and data preprocessing methods.The specific research content of this paper is as follows:For high-resolution image classification,this paper proposes a classification model based on the enhanced Deep Lab V3 network,which uses the characteristics of different receptive fields of convolution kernels of different sizes to construct the R-MCN(Resnet Multi Convolutional Network)module,and replace the conventional convolution with horizontal and vertical separate convolution kernels to reduce the amount of model calculation parameters,extracts the multi-scale semantic information of the feature map,and then uses a learnable upsampling method to recover the detail loss caused by the downsampling or convolution operation.and finally use the improved loss function Mloss.By comparing with several commonly used deep learning image classification algorithms,the effectiveness of the method proposed in this article is verified.For remote sensing object detection,current convolutional neural networks are difficult to automatically fuse the optimal feature information and are sensitive to scale variation of different objects.To address that this paper proposes an adaptive attention fusion mechanism convolutional network model.The proposed method is based on the backbone network of Efficient Det.Firstly,according to the characteristics of object distribution of datasets,the stitcher-image stitching scheme is used to potentially introduce objects with majority sizes.This process can effectively balance the datasets and mitigate the scale variation issue.Then an adaptive attention fusion mechanism(AAFM)is built,which concentrates the spatial and channel information of feature maps by convolution and different pooling operations and perform the fusion of optimal proportion using the fusion factor learned from the network.Finally,the CIo U Loss is used to regress the bounding box that is closer to the ground truth.The experimental results on the optical image dataset DIOR demonstrate that,the proposed method can effectively improve the detection accuracy and has stronger robustness compared with the state of the art detection algorithms such as SSD,YOLOv4,and Efficient Det.