CN-115510660-B - Infrared dim target detection method based on deep sparse low-rank neural network

Abstract

An infrared weak and small target detection method based on a deep sparse low-rank neural network belongs to the field of infrared data processing in remote sensing digital image processing. Because the infrared weak and small target is small in size and low in brightness, the infrared weak and small target is difficult to detect from an image, the invention provides a method for improving the performance of an infrared weak and small target detection algorithm in the aspect of small target detection and solves the problem that the target detection result under the clutter background is inaccurate. The method comprises the steps of dividing an original infrared image into a series of infrared image blocks by utilizing a sliding window, establishing a target detection model based on target sparse representation and background low-rank constraint, inputting the infrared image blocks, solving various variables of the target detection model by utilizing an alternating direction multiplier method, expanding the proposed model into a convolutional neural network, continuously updating relevant parameters in the model, reconstructing a target detection result in the obtained infrared image blocks, and outputting the target detection result of the infrared image. According to the invention, good detection results can be obtained for infrared targets with different attributes under different background environments.

Inventors

• HU YUE
• ZHOU XINYU
• ZHANG YE

Assignees

• 哈尔滨工业大学

Dates

Publication Date

20260508

Application Date

20220930

Claims (4)

1. 1. The method for detecting the infrared weak and small target based on the deep sparse low-rank neural network is characterized by comprising the following steps of: Step one, inputting a piece of space with the size of Infrared image of (a) Designing a sliding window According to a fixed step length From infrared images Moving from the upper left corner to the lower right corner, and extracting an infrared image The local image blocks in the image are extracted to obtain all local image blocks, and the local image blocks are vectorized to form a space with the size of Is an infrared block image of (a) ; Step two, utilizing low rank and sparse characteristics of the background and the target in the infrared block image to establish a convex optimization equation to detect the infrared weak and small target, and based on the background consistency assumption, pushing out the background of the infrared block image Is a low-rank matrix, and simultaneously utilizes the difference between the target and the background to deduce the target infrared block image Is a sparse matrix; step three, enhancing a target sparse constraint term in the convex optimization equation by using sparse learning, and reconstructing the established convex optimization problem by using an augmented Lagrangian algorithm Solution to (c) introducing auxiliary variables On the basis, the target infrared block images are respectively and iteratively solved by using an alternate direction multiplier method Infrared block background image Auxiliary variable Lagrangian multiplier ; Expanding the established sparse target detection model into a convolutional neural network, and respectively designing a sparse prior layer, a low-rank prior layer, a reconstruction layer and a multiplier updating layer to update and solve parameters involved in the process of solving each variable and auxiliary variables; step five, judging the infrared block image And iteratively outputting a target infrared block image Infrared block background image If yes, outputting the target infrared block image And infrared block background image If not, returning to the third step to continue the iteration loop; Step six, outputting the target infrared block image And infrared block background image The method comprises the steps of firstly, referring to a vectorization method in the first step, recovering the vectorization method into local image blocks, splicing the local image blocks according to the step length of a sliding window, taking the median value of pixel values in the local image blocks corresponding to pixels at the same position in an original image as a final gray value in an output image, and thus obtaining an infrared dim target detection result.
2. 2. The method for detecting the infrared dim target based on the deep sparse low rank neural network according to claim 1, wherein the specific process of the second step is as follows: Step two, the extracted local image block is related with the surrounding image blocks due to the influence of heat radiation and diffraction, and shows non-local autocorrelation characteristic Is a low rank matrix, and on the contrary, the pixel intensity of the weak and small target is usually greatly different from the surrounding pixel intensity, and the infrared block image of the target is usually obtained Is a sparse matrix, thus, the infrared small target detection problem can be expressed as (1) Wherein, the Is an image of an infrared block of light, Is the noise of the sound and, Is a regularization parameter which is a function of the data, Representative of The norm of the sample is calculated, Representing the rank of the matrix; Step two, step two, The solution of the norm is an NP-hard problem, which cannot be solved directly, and therefore, uses Norm substitution Norm and solving the rank of the matrix with the kernel norm, equation (1) can be rewritten as (2) Wherein, the Representing the norm of the kernel and, Representative of The norm of the sample is calculated, Represents the normal number of Frobrnius and, Is a constant close to 0.
3. 3. The method for detecting the infrared dim target based on the deep sparse low rank neural network according to claim 2, wherein the specific process of the third step is as follows: Step three, the object detected by the formula (2) is generally affected by image background and noise, so that the sparsity of the detected object is enhanced by sparse conversion, then the formula (2) can be expressed as (3) Wherein, the Is a sparse transition; step three, optimizing the formula (3) by using an augmented Lagrangian multiplier method to obtain the following expression form (4) Wherein, the Is a penalty parameter which is a function of the penalty parameter, Is the lagrange multiplier and is a function of the lagrange, And Is a regularization parameter; Step three, respectively solving the formula (4) by using an alternate direction multiplier method And The following is shown (5) Wherein, the , Is the update rate; Step III, IV, for The solution of (a) adopts a singular value threshold algorithm, and is as follows (6) Wherein, the , And Is to The first of the results obtained after singular value decomposition The number of the values to be used in the process, Representative maximum value is obtained; represents the first Generated by multiple iterations Is a value of (2); step III, five, variable Is a linear inverse problem, which is solved using an iterative shrinkage threshold algorithm, in particular, by alternately solving the auxiliary variables using an iterative shrinkage threshold algorithm And The following is shown (7) (8) Wherein the method comprises the steps of Is the update rate of the data to be updated, The solution can be performed by using a soft threshold method, i.e 。
4. 4. The method for detecting the infrared dim target based on the deep sparse low rank neural network according to claim 3, wherein the specific process of the fourth step is as follows: step four, designing a reconstruction layer The reconstruction result of the t-th iteration is related to the output of the t-1 st iteration, i.e., according to equation (7) 、 And In the first iteration, the first time of the iteration, 、 And Is arranged as Super-parameters May be updated in each iteration; Initializing the value to 0.1; step four, designing a sparse prior layer The layer enhances the target according to the continuous update of the formula (8) A 6-layer convolutional neural network is designed, and a complete sparse transformation matrix is obtained through the study, training and verification of hundreds of groups of training data In order to expand the network capacity, in equation (8) And Respectively learning by using a three-layer convolutional neural network in the layer, and obtaining And ; Step four, three, design low rank prior layer According to formula (6), Solving by singular value decomposition with soft thresholding scheme, deep sparse low rank neural network where the thresholding is learned to obtain the best value, as follows (9) Fourth step, design multiplier update layer The Brownian multiplier is updated to formula (5) at this layer; The super-parameters trained in the deep sparse low-rank neural network are initialized to 0.1; Step IV, five, parameters And And And continuously learning and updating in the designed network, sharing parameters among the networks, and finally outputting a complete target detection network.

Description

Infrared dim target detection method based on deep sparse low-rank neural network Technical Field The invention belongs to the field of infrared data processing in remote sensing digital image processing, and particularly relates to an algorithm for detecting a weak and small target in an infrared image based on a deep sparse low-rank neural network. Background Infrared small target detection with night detection capability is of great importance in target search and tracking systems, which have been widely used in the field of military surveillance and precision guided weapons. Presently recognized definition of infrared dim objects is that the infrared object size is in the range of 2 x 2 and 9 x 9, or the proportion of the whole image is less than 0.15%. The infrared search and tracking system relies on accurate detection results of the target detection method. However, since the infrared weak target is often characterized by small volume and low brightness, the infrared weak target is not generally characterized by space. And the target to be detected is often in a complex background or is affected by sea clutter and cloud clutter, so that noise and background are often erroneously detected as targets. In recent years, many scholars have been working on improving the performance of the infrared dim target detection method and various detection algorithms have been proposed for this task. Generally, the proposed infrared dim target detection method can be broadly divided into two categories, namely a pre-detection tracking algorithm and a pre-tracking detection algorithm. The method based on pre-detection tracking aims at tracking the infrared weak and small target in the multi-frame image on the premise of being based on the consistent track assumption. Many pre-detection tracking algorithms are proposed based on 3D matched filtering and maximum likelihood estimation calculations. Such algorithms typically require processing hundreds of frames of data and thus typically have a high temporal complexity. Models based on pre-tracking detection methods aim to suppress the background and highlight small targets. The method is divided into 4 methods, namely a background space consistency-based method, a neural network-based learning model, a target significance-based method and an infrared block image-based method. These algorithms all require calculation of the characteristics of the target and the background, and detection of the image by means of target enhancement and background suppression. However, because the weak target does not have specific spatial characteristics, the detection method before tracking is often interfered by background and noise. Disclosure of Invention The invention provides an infrared dim target detection method based on a deep sparse low-rank neural network for solving the problem of accurately detecting a dim target in an infrared image under different environmental backgrounds. And detecting an infrared small target through the constructed sparse infrared image block model, iteratively updating variables in the model by using an alternate direction multiplier method, and expanding the constructed model into a convolutional neural network to update model parameters so as to obtain a final target detection result. In order to solve the problems, the invention adopts the technical scheme that the content of the authorization document is as follows: The method has the technical key points that an infrared image to be detected is input, and a sliding window moving from top left to bottom right is arranged in the infrared image to extract infrared local image blocks. These local tiles are vectorized to form a matrix of tiles. Because background information generally has low rank characteristics and weak targets have sparse characteristics in each image block, we construct an optimization equation to solve the infrared weak target detection problem. Furthermore, the edges and corner points are suppressed using a learning sparse transform in sparse regularization terms. The proposed optimization model can be solved efficiently with an alternating direction multiplier algorithm. And finally, expanding the iteration step of the proposed weak and small target detection model into a depth network. The proposed deep sparse low-rank neural network comprises four layers, namely a sparse prior layer, a low-rank prior layer, a reconstruction layer and a multiplier updating layer. Training the learning sparse transformation by designing a plurality of convolutional neural network layers. Parameters in the network are trained and shared between the different layers. The invention adopts the technical scheme for solving the problem of infrared weak and small target detection that: Step one, an infrared image f D with a spatial dimension of MxN is input. A sliding window w is designed, moving from the upper left corner to the lower right corner of the infrared image f D according to a fixed step s,