| The problem of minimal target detection in remote sensing images has always been one of the difficulties in the research of target detection.Due to the small imaging area of the minimal target(equal to or less than 20×20 pixels)and the small amount of information carried,the feature information of the texture edge or shape of the target is very short,and it is difficult to achieve effective feature extraction and fusion.The traditional target detection technology has poor effect on remote sensing images.Although the proposed convolutional neural network simplifies the detection process and improves the detection accuracy,the detection performance on remote sensing images is still not ideal.In this paper,the minimal target detection algorithm for remote sensing images is studied.The main work is as follows:(1)A new Remote Sensing Image Minimal target dataset,RS-M,is proposed to solve the problem of the lack of open Remote Sensing image Minimal target dataset.Most of the targets are 20 × 20 pixels or less.In addition,a variety of data enhancement methods,including Cut Out method,Mix Up method,Cut Mix method,Mosaic method and Bilateral Blurring method,are used to expand the data set,and the enhanced ERS-M data set is obtained.The number of images in the dataset increases by about 4 times and the number of minimal objects increases by about 10 times.(2)A Strengthened Feature Pyramid Network(SFPN)is proposed to solve the problem of detection of extremely small objects.A bottom-up path is added to the FPN for Feature fusion.It can realize fusion of features of different depth of minimal target.The proposed SFPN is verified on different backbone networks,and the results show that SFPN improves the detection accuracy of minimal targets.(3)Aiming at the poor performance of the general detection model in minimal target detection,a network SAD-CSPDarknet for minimal target detection is proposed.Firstly,attention mechanism and spatial pyramid pooling structure are added to the backbone network to effectively extract important features,adjust the receptive field of the network,and enhance the ability of the network to extract minimal targets.Secondly,attention mechanism is added into the feature fusion network to improve the feature fusion capability of the network.Thirdly,the convolutions of two different channel numbers are added to the detection sub-network to compress the number of network parameters,which is beneficial to further improve the detection accuracy.Finally,the proposed network is compared with the detection network of YOLO series on the ERS-M data set,and the experimental results show that the proposed SAD-CSPDarknet can improve the detection accuracy of minimal targets. |
