Research On Infrared Dim Small Targets Detection Algorithm Based On Image Enhancement And Semantic Segmentation Network

Infrared small target detection has important application value in military early warning.Weak and small target detection has problems such as long imaging distance,small area,and fewer pixels on the imaging plane of the infrared image.Hand-designed target features are difficult to correctly distinguish targets,backgrounds and noises in complex background images,and there are many existing algorithms It has high computational complexity and does not meet military real-time requirements.High-precision detection is still a challenging task.Therefore,it is necessary to conduct better algorithm exploration.The main contributions of this paper are as follows:(1)In order to highlight the target and suppress the background,an image enhancement algorithm that can effectively suppress the background and improve the saliency of the target is first proposed,and compared with the existing image enhancement algorithm,the effectiveness of this algorithm is proved.(2)The background of dim targets is low-rank and sparse.Based on this,this paper proposes a dim target detection algorithm based on sparse filtering.First cut the image into image blocks,treat each image block as uncorrelated data samples,and then vectorize,and add the ideal point target matrix as the sentinel matrix to the data samples.The sentinel matrix also needs to be vectorized;then All vectors use sparse filtering to map from low-dimensional space to high-dimensional space.The sentinel vector is used to control the iterative stop of sparse filtering.The value of the vector is used to determine the similarity between the original data block and the target in the feature matrix.Then,in order to solve the problem that sparse filtering is sensitive to image edges,the combined algorithm of gradient matrix singular value decomposition and structure vector tracing is used to filter out the image edges and retain the target.The main idea of the algorithm is to transform the image to obtain more features,and the feasibility of the algorithm is proved through experiments.(3)A method for weak and small target detection based on semantic segmentation network is proposed,which has high detection accuracy and real-time detection capability.This paper first proposes the FCN-Based Network,and proves the feasibility of the semantic segmentation network to detect small targets through training and generalization results;then proposes an improved version of the STDNN network,taking into account the multi-scale problem of the target,adding a hole convolution to the network,Pyramid network and other structures;in order to solve the problem of the serious imbalance between the target and the background,the Tversky loss function is adopted;in order to improve the accuracy of network detection,the structure of cross-connection and residual connection is added.The trained network can accept images of any size as input and quickly obtain background and target segmentation images.Finally,considering the influence of noise,the connection method is used to filter the misjudged noise points on the segmentation map to improve the detection rate.The network training data set is obtained by dividing the infrared image collected under the clear background into small blocks and adding different signal-to-noise ratios and multi-scale simulation targets to it.It is proved by testing and comparison experiments on the simulation data set and the real data set.The network model can quickly and accurately segment the target and background in the image,and the detection effect is better than all comparison methods.