Research On Infrared Small Target Segmentation And Detection Method Based On Deep Learning

In recent years,infrared small target detection tasks have received a lot of attention,and are widely used in maritime monitoring and early warning systems.The traditional infrared image small target detection method relies heavily on the setting of manual features,which leads to the instability of the model and it is difficult to adapt to the changes of the scene;and the direct detection method using deep learning is also difficult to detect the real small target due to the scale problem.In order to better realize the detection of infrared small target images,this thesis proposes an infrared small target segmentation detection network GSTD-Net based on texture enhancement,and based on it,a network loss function of strong and weak supervision is constructed.The specific method as follows:(1)GSTD-Net is based on the U-Net backbone structure and is also divided into two parts:encoding and decoding.Considering the characteristics of small target pixel information in infrared small target images,it is not necessary to add skip connections to link encoding features and decoding features,and a simple serial structure can already meet the small target task.Different from the ordinary U-Net network,GSTD supplements the image prior information in time in the early coding process to retain the key target features in the image as much as possible,so that it is not lost in the layer-by-layer network.The decoding network is mainly composed of a texture enhancement layer—Gabor convolutional neural network layer.The network layer is designed by imitating Gabor filtering.The unique biological characteristics of Gabor filtering are used to enhance the ability of the network to extract features,which can be fully and effectively extracted to the target internal structure.features,which are robust to changing scenes.(2)In addition,considering that small targets are small in size and easily submerged in clutter,a network loss function with strong and weak supervision is constructed on the basis of GSTD-Net.Using KL loss as a weaker supervision,and adopting the strategy of regression prediction probability distribution to achieve KL loss,can predict the uncertainty in the network more flexibly.Using this combination of strong and weak supervision can more comprehensively supervise all feature information of the network,approach the network in a more efficient and accurate way,and further improve the accuracy of network detection.The infrared small target detection method proposed in this thesis has been experimented on several public infrared small target datasets and our own created dataset NUCC-SIRST.The experimental results show that,compared with the traditional small target detection method and other deep learning detection methods,the detection algorithm proposed in this thesis has greatly improved the detection probability.The effectiveness and superiority of this algorithm are verified.