Research On Infrared Small Target Detection Algorithm Based On Deep Learning

Infrared imaging system,which works by passively receiving infrared electromagnetic wave radiation from the object,has the advantages of all-weather work,strong anti-interference ability,and long detection distance.Because of these advantages,target detection and recognition technology based on infrared imaging has important application value in both military and civilian fields.Compared with visible images,infrared images have the disadvantage of low resolution and signal-to-clutter ratio of target,which leads to the following difficulties in the detection of infrared small targets:（1）There are many complex backgrounds,such as clouds and sea surface,resulting in low signal-to-clutter ratio of target;（2）The target is usually shown as a dim target in the image and texture feature is lacking when infrared imaging distance is relatively long;（3）Because the target is small and sparse,there is a serious class imbalance problem between target and background.The research is performed in this paper to solve the above problems,and the main work contents are as follows:（1）Considering infrared small target detection as a semantic segmentation problem,an end-to-end infrared small target segmentation model integrating multi-scale fractal attention is designed.Firstly,based on the analysis of multi-scale fractal features suitable for dim target detection in infrared images,its accelerated calculation process using deep learning operator is presented.Secondly,the metric based on convolutional neural network is designed to obtain the target significance distribution map.An attention module based on multi-scale fractal feature is proposed combining the feature pyramid attention module and pyramid pooling down-sampling module.When it is embedded into the semantic segmentation model of infrared targets,the asymmetric context fusion mechanism is used to improve the fusion performance of shallow features and deep features,and an asymmetric pyramid non-local module is used to obtain global attention to improve the detection performance of infrared small targets.Finally,a single frame infrared small target（SIRST）detection dataset is used to verify the performance of the proposed algorithm.The intersection over union（Io U）and the normalized intersection over union（n Io U）of the proposed model reach 77.4%and76.1%,respectively,which is better than the performance of the known methods.Meanwhile,the effectiveness of the proposed model is further verified by the migration experiment.Due to combining prior knowledge of traditional methods and feature learning ability of deep learning methods effectively,the proposed model is suitable for infrared small target detection in complex scenarios.（2）An infrared small target real-time detection algorithm based on improved YOLOv5s is presented,which is improved from three aspects of feature extraction,feature fusion and prediction output.1)In the feature extraction stage,the problem of feature loss caused by the down-sampling stage of YOLOv5s is first analyzed,and SPD-Conv module is adopted to replace the down-sampling module in YOLOv5s,so as to retain the features of the infrared small target to the maximum extent.The feature extraction capability is then improved using channel attention and spatial attention modules.An improved atrous spatial pyramid pooling module is finally designed to concatenate convolution kernel with different receptive fields through different atrous rates to obtain multi-scale features;2)In the feature fusion stage,the top-down attentional module is introduced to embed deep semantic features into shallow spatial features,which improves expression ability of shallow features and the effect of feature fusion;3)In the prediction stage,the prediction layer for large targets in the network and the related feature extraction layers and feature fusion layers are reduced,which greatly reduces the model size and improves detection speed.Finally,the effectiveness of each module is verified by ablation experiments.The experimental results show that,m AP_0.5 of the improved model on SIRST dataset reaches 95.4%,which is 2.3%higher than that of YOLOv5s.The model size is greatly reduced to only 27.1%of the original YOLOv5s,and the inference speed reaches 172 frames per second（FPS）,which is conducive to the actual deployment and application.The effectiveness of the improved model is also verified by the migration experiment on Infrared-PV dataset.The proposed model has high detection rate and low false alarm rate,which is suitable for small target detection in infrared images.