Research On Infrared Small Target Detection Algorithm Based On Low-rank Sparse Recovery Theory

Infrared image processing technology is an important research field of today's digital image processing technology.It plays an important role in military,medical,industrial and other fields.For example,the precise positioning of small infrared targets in military guidance can help predict the situation of war and obtain intelligence in a timely manner.Therefore,infrared target detection technology is a subject worthy of in-depth study in infrared image processing.Due to the complex background structure in the infrared image,the target is usually a few to a dozen non-zero pixels,resulting in low image noise,so how to extract small targets from the complex background with low false alarm rate and high detection rate is a very difficult question.Inspired by the study of traditional classical and popular target detection algorithms,this paper puts forward its own idea,which combines image spatial domain,frequency domain and low rank sparse theory to apply to infrared target detection,promotes each other,makes various algorithms complement each other,and improves the robustness of detection effect.The main work is as follows:(1)In view of the small target size is unknown,different Gaussian kernel templates in the airspace have different smoothing effects on the image,and the processing of the image by a single template is easy to cause missed detection.A low-rank sparse recovery algorithm based on multi-scale Gaussian space is proposed.Then,the target saliency map is obtained by making a difference with the original image.First,convolve the Gaussian template with different scales and the image to obtain a smooth image after suppressing the target at three different scales;further use the accelerated near-end gradient method to perform low-rank sparse recovery on the smoothed image to ensure that each scale is obtained The low-rank matrix does not contain target information,and the maximum value of the low-rank matrix of each scale is taken as the final background image;then the original image and the background image are subtracted to obtain the target saliency map.Finally,the mean and variance of the image are used for the saliency The graph is thresholded to obtain the final target detection result.The experimental results show that the difference between the original image and the background image obtained by fusing different scales improves the contrast between the target and the background and also ensures the integrity of the target in the saliency map.It has higher detection rate and Low false alarm rate in comparison with other comparison algorithms.(2)Aiming at the small target in infrared image with sparseness and high frequency,the background has low rank and low frequency,this paper proposes an infrared small target detection algorithm based on low rank sparse recovery and image block discrete cosine transform.Further suppress the background clutter in the low rank sparse recovery algorithm.Firstly,the low-rank sparse restoration is performed on the image using the alternating direction method to obtain the sparse target saliency map;then the saliency map is divided into two scale non-overlapping blocks,and the discrete cosine transform is performed on each block,set the DC coefficient to further suppress the remaining clutter background in the saliency map,perform inverse transform to obtain the processing results of two transform domains with different scales,normalize the two inversely transformed images to perform a dot product operation,and further enhance the target and background To obtain the final target saliency map;and finally perform threshold segmentation to achieve target detection.Experimental results show that the algorithm can effectively suppress various complex backgrounds and reduce the false alarm rate.(3)Aiming at the problem of large amount of matrix data caused by the classic image blocking model,a low-rank sparse recovery algorithm based on image blocking and compression domain is proposed,and different-scale blocking schemes are proposed for original images of different sizes.Firstly,according to the proposed image partitioning scheme,the original image to be processed is set with the corresponding sliding window and the step size is overlapped,and each block is drawn into a column vector and a large matrix according to the order of the block,to further enhance Correlation between background pixels;then use a random measurement operator to sample the large matrix after column vectorization to obtain a one-dimensional vector that is much smaller than the size of the large matrix to reduce the amount of data that needs to be processed;further The low-rank sparse restoration is performed on the dimension vector to obtain the sparse block image,and finally the sparse block image is reconstructed to obtain the final target detection image.By comparing experimental simulation with other algorithms,it is further proved that the algorithm can be adapted to target detection in a variety of complex backgrounds,and has a higher detection rate and a lower false alarm rate.