Research On Critical Algorithms Of Infrared Small Target Detection Based On Deep Learning

The core idea of the infrared small target detection task is to find the location of the target in a set of infrared image sequences,so as to carry out subsequent target tracking tasks.In previous studies,deep learning neural network models represented by Faster R-CNN and YOLO have developed rapidly.At the same time,some algorithms based on local filter have also been shown to be very successful in improving the discriminatability of the target detection process.However,these algorithms build patches by traversing local images,ignoring the correlation between different image areas.This will cause some texture information of the target to be ignored,eventually making the algorithm unable to accurately identify any form of target.Non-local detection algorithms use global image information for processing and training,which can effectively improve the robustness of the detection model,but this type of algorithm has the disadvantages of large amount of computation and slow speed.In order to improve the generalization of the infrared small target detection algorithm in different scenarios,two algorithms are proposed in this thesis,and the main research results are as follows:(1)Aiming at the problem that the traditional local infrared small target detection algorithm ignores the correlation between different image regions,this thesis proposes a new target detection algorithm based on pyramid co-occurrence feature architecture.The newly proposed detection framework can decompose the infrared small target detection problem: Firstly,through the pyramid feature architecture,the input infrared image is layered and the corresponding co-occurrence matrix is extracted,and then these co-occurrence matrices are fused to better summarize the non-local spatial sparsity limitations.On the other hand,co-occurrence matrices preserve semantic contextual information about the target and surrounding region,thus avoiding deviations from the approximation of low-rank matrices.Finally,by combining the adaptive diamond-shaped structural elements for morphological operations,we can get the final infrared small target detection results.(2)For existing deep learning saliency target detection architectures,while models are able to capture richer image local and global information,they often require very large and diverse training datasets to train more accurate and robust models.For infrared images,although the background contains a variety of scenes,they are all affected by noise and appear black and gray in the subjective vision,and the overall class distinction is not obvious.Aiming at this problem,we propose a deep learning target detection model based on co-occurrence graph segmentation neural network.The model uses co-occurrence filter as a new convolutional layer in the training process and then combines the idea of residuals to design the co-occurrence residual block.Then,it is incorporated into the deep learning graph segmentation model to improve the recognizability of the infrared image in the training process.On the other hand,the reverse attention mechanism is introduced into the model.By generating a relatively rough mask map,and then removing the salient regions of the current prediction in the side input features,the model can supplement the corresponding details.Finally,the output maps of each layer are fused to obtain the final infrared small target detection result.In order to explore the performance of the infrared small target detection algorithms proposed in this thesis,this thesis conducts a comparison experiment based on the real collected infrared sequence images,and thirteen advanced infrared small target detection algorithms are involved in the comparison.Through the subjective evaluation of the detection results and the objective evaluation indexes,the infrared small target detection algorithms proposed in this thesis can show good performance in various complex infrared images,and ensure the generalization of the algorithms.