| Infrared small target detection is a technology for locating and segmenting small targets in infrared images,which has important practical value in precision guidance,weapon manufacturing,monitoring,and early warning.However,because the radiation intensity of small targets is weakened by atmospheric radiation and environmental radiation,coupled with long imaging distance and complex imaging environment,small targets in infrared images are small in size,weak in signal,and lack shape and texture.Therefore,infrared small target detection is a very challenging task.According to the characteristics of infrared small targets,this thesis models the detection task as a segmentation task and researches infrared single-frame small target detection algorithm based on U-Net architecture.Through an in-depth analysis of the characteristics and detection difficulties of small infrared targets,this thesis proposes a U-Net-based infrared small target detection algorithm,which has the advantages of a high detection rate and low false alarm rate.The main contributions and innovations of this thesis are summarized as follows.To solve the problem of insufficient detection accuracy of small infrared targets,this thesis deeply analyzes the size of the image,the size of the infrared small target and its background category,and proposes an infrared small target detection algorithm based on the Cross-layer Context Information Aggregation module(CCIA).In this algorithm,a random scale input module is designed to strengthen the model’s ability to handle multi-scale objects.Then the thesis proposes a dual attention context information aggregation module,which realizes the efficient fusion of deep and shallow network information from coarse to fine.Specifically,the model first uses channel attention to filter deep and shallow features separately,and then uses shallow detail features to perform attention modulation on deep semantic features.Finally,the shallow feature map is fused with the modulated features after being processed by the bottom-up point-wise attention module.The whole module realizes the aggregation of context information in channel dimension and space dimension,and fully retains and highlights the characteristics of small infrared targets.At the same time,this thesis redesigns the downsampling scheme of U-Net and uses the residual block as the feature extraction module in the up and down sampling process,which effectively solves the problem of the easy loss of small target information while alleviating the phenomenon of gradient disappearance.The experimental results show that the CCIA algorithm has higher detection accuracy for small targets compared with U-Net.CCIA’s Io U and n Io U are increased by0.03 and 0.023 respectively,the detection rate is increased by 8.22%,the false alarm rate is reduced by 59.22%,the network model complexity is reduced by 7%,and the running time is 0.155 seconds,which is only 0.011 seconds slower than U-Net,indicating that the CCIA algorithm has good real-time performance.Due to the small number of pixels occupied and the lack of clear shape and texture,small infrared targets are susceptible to clutter and noise interference in complex environments,resulting in serious problems of missed detection and false detection in target detection tasks.Therefore,to further reduce the false alarm rate,based on the CCIA algorithm,this thesis proposes a non-local dependency-based infrared small target detection algorithm(NLDCCIA).Considering that the convolution operation and the pooling operation are both local operations and cannot pay attention to the global information,a cross-layer non-local feature fusion module is designed inspired by the idea of non-local attention.This thesis combines it with a non-local attention module and adds it to the CCIA network.Firstly,the non-local attention module can acquire long-range dependencies and better distinguish small objects from false alarms.Secondly,the cross-layer non-local feature fusion module calculates nonlocal dependencies after down-sampling shallow features,which effectively reduces the amount of computation.By incorporating deep semantic feature dependencies,it can obtain richer global dependency information and further reduce the false alarm rate.The experimental results show that the Io U and n Io U of the NLD-CCIA algorithm are increased by 0.008 and 0.006 respectively compared with the CCIA algorithm.The detection rate is increased by 0.02%,the false alarm rate is reduced by 2.78%,the network model complexity is only increased by 0.03 M,and the running time is only slowed down by 0.056 seconds.It shows that the NLD-CCIA algorithm can improve the detection accuracy while reducing the false alarm rate,and has a lightweight network structure,which can meet the real-time requirements of practical applications. |
