Patch-based methods improve the performaInce of inffrared small tSmall-Target dDetection, transforming the detection problem into a Low-Rank Sparse Decomposition (LRSD) problem. However, two challenges hinder the success of these methods: (1) The interference from strong edges of the background, and (2) the time-consuming nature of solving the model. To tackle these two challenges, we propose a novel infrared small-target detection method using a Background-Suppression Proximal Gradient (BSPG) and GPU parallelism. We first propose a new continuation strategy to suppress the strong edges. This strategy enables the model to simultaneously consider heterogeneous components while dealing with low-rank backgrounds. Then, the Approximate Partial Singular Value Decomposition (APSVD) is presented to accelerate the solution of the LRSD problem and furthe (ISTD) is an important component of infrared search and tracking, aiming to exploit the thermal radiation difference between a target and its background to achieve long-range target detection. According to the definition by the Society of Photo-Optical Instrumentation Engineers (SPIE), small tar
improvge the solution accuracy. Finally, we implement our method on GPU using multi-threaded parallelism, in order to further enhance the computational efficiency of the model. The experimental results demonstrate that our method outperforms existing advanced methods, in terms of detection accuracy and execution times typically refers to objects in a 256 × 256 image with an area of fewer than 80 pixels, accounting for approximately 0.12% of the total image area.
HVS-based methods detect small targets by utilizing the contrast differences between the target region and its surrounding background. These methods can be categorized based on the type of information they use: grey scale information, gradient information, and a combination of both greyscale and gradient information. Local Contrast Measure (LCM) [1] proposes a novel method for detecting small targets by leveraging grey scale contrast. This method uses a contrast mechanism designed to enhance small targets while effectively suppressing the background noise. Based on the improvement of the LCM algorithm, Relative Local Contrast Measure (RLCM) [8], Multiscale Patch-based Contrast Measure (MPCM) [9], Weighted Local Difference Measure (WLDM) [27] and other methods were proposed. Gradient-based contrast methods use first-order or second-order derivatives of the image to extract gradient information. They then utilize this information to design a gradient difference measure that effectively discriminates between small targets and the surrounding background. Building on this concept, Derivative Entropy-based Contrast Measure (DECM) [28] and Local Contrast-Weighted Multidirectional Derivative (LCWMD) [29] propose the use of multidirectional derivative to incorporate more gradient information. In addition, Local Intensity and Gradient (LIG) [30], and Gradient-Intensity Joint Saliency Measure (GISM) [31] fuse gradient and intensity information to further highlight small targets. Although HVS-based methods can be effective in many scenarios, they are susceptible to missed detections and false positives in images characterized by low signal-to-clutter ratios and high-intensity backgrounds.
In recent years, there has been a significant research focus on deep learning-based methods for infrared small target detection, which seek to achieve high-accuracy detection rates. These deep learning models are trained to discern features within infrared images using vast datasets, thereby enhancing their detection capabilities. To address the problem that infrared small target features are easily lost in deep neural networks, Attention Local Contrast Network (ALCNet) [32] proposes asymmetric contextual modulation to interact with the feature information between the high and low levels. Dense Nested Attention Network (DNANet) [15] adequately fuses feature information through densely nested interaction modules to maintain small targets in deep layers. Miss Detection vs. False Alarm (MDvsFA) [33] proposes dual generative adversarial network models, trained inversely to decompose the detection challenge into sub-problems, aiming to strike a balance between miss detections and false alarms. While publicly available datasets have advanced deep learning for infrared small target detection, the scant features of small targets and the dependency on training samples limit the applicability of the model in varied real-world scenarios.
A significant amount of research has been conducted to improve the detection ability of IPI [17]. On one hand, some methods have used prior constraints, including ColumnWeighted IPI (WIPI) [18], Non-negative IPI with Partial Sum (NIPPS) [20], and Re-Weighted IPI (ReWIPI) [21]. On the other hand, some studies have identified limitations in the nuclear norm and L1 norm and, so, alternative norms to achieve improved target representation and background suppression have been proposed; for example, Non-convex Rank Approximation Minimization (NRAM) [22] and Non-convex Optimization with Lp norm Constraint (NOLC) [23] introduce non-convex matrix rank approximation coupled with L2,1 norm and Lp norm regularization, while Total Variation Weighted Low-Rank (TVWLR) [24], Kernel Robust Principal Component Analysis (KRPCA) [25] introduce total variation regularization, High Local Variance (HLV) [26] method present LV* norm to constrain the background’s local variance. Patch-based methods mainly consider the low-rank nature of the background, affecting their performance in the presence of strong edges. However, our method pays additional attention to heterogeneous background suppression in low-rank constraints, to avoid this problem.
Acceleration strategies for patch-based methods can be categorized into algorithm-level and hardware-level acceleration. The first category mainly relies on the strategy of reducing the number of iterations. Self-Regularized Weighted Sparse (SRWS) [34] and NOLC [23] improve the iteration termination condition for acceleration but still suffer from the time consumption associated with decomposing large matrices. The other category (i.e., hardware acceleration) relies on the use of computationally powerful hardware and efficient parallelization strategies. In Ref [35], the researchers proposed Separable Convolutional Templates (SCT); however, this method has poor performance under complex backgrounds. In addition, extending the patch model to tensor space can also achieve acceleration [36][37][38][39][40][41]. Representative methods in this direction include Re-weighted Infrared Patch-Tensor (RIPT) [36], LogTFNN [39] and the Pareto Frontier Algorithm (PFA) [37]. However, unfolding the tensor into a two-dimensional matrix before decomposition increases the algorithm’s complexity. Partial Sum of the Tensor Nuclear Norm (PSTNN) [38] and Self-Adaptive and Non-Local Patch-Tensor Model (ANLPT) [42] utilize the t-SVD speed up tensor decomposition with t-SVD. However, these methods are limited by the complexity of finding the applicable constrained kernel norm. Our work investigates accelerated patch-based methods at both the algorithmic and hardware levels.