PaperlayerView exact revision on arXiv

Source-linked AI summary

Reweighted Infrared Patch-Tensor Model With Both Non-Local and Local Priors for Single-Frame Small Target Detection

Yimian Dai, Yiquan Wu

arXiv:1703.09157v1cs.CV

TL;DR

Existing infrared small-target detectors struggle with heterogeneous backgrounds because strong edges interfere and local or non-local priors are used incompletely. The paper proposes RIPT, combining an infrared patch-tensor model with local-structure-adaptive and sparsity-enhancing reweighting. Experiments report robust background suppression and clearer target separation, especially for dim targets and heavy clutter.

  • Problem

    Existing methods can fail on heterogeneous backgrounds because local-prior methods suffer rare-structure effects and non-local-prior methods retain salient edge residuals.

  • Method

    RIPT represents images with an infrared patch-tensor and performs reweighted robust tensor recovery using complementary non-local and local structure priors.

  • Results

    The proposed model is reported to achieve robust performance across varied scenes and clearer background-suppressed target images than the tested state-of-the-art methods.

  • Takeaways & Limitations

    Combining local structure with non-local redundancy supports target-background separation when dim targets and strong clutter challenge either prior alone.

Abstract

from arXiv · show

Many state-of-the-art methods have been proposed for infrared small target detection. They work well on the images with homogeneous backgrounds and high-contrast targets. However, when facing highly heterogeneous backgrounds, they would not perform very well, mainly due to: 1) the existence of strong edges and other interfering components, 2) not utilizing the priors fully. Inspired by this, we propose a novel method to exploit both local and non-local priors simultaneously. Firstly, we employ a new infrared patch-tensor (IPT) model to represent the image and preserve its spatial correlations. Exploiting the target sparse prior and background non-local self-correlation prior, the target-background separation is modeled as a robust low-rank tensor recovery problem. Moreover, with the help of the structure tensor and reweighted idea, we design an entry-wise local-structure-adaptive and sparsity enhancing weight to replace the globally constant weighting parameter. The decomposition could be achieved via the element-wise reweighted higher-order robust principal component analysis with an additional convergence condition according to the practical situation of target detection. Extensive experiments demonstrate that our model outperforms the other state-of-the-arts, in particular for the images with very dim targets and heavy clutters.

I. INTRODUCTION

Single-frame infrared small target detection is needed because changing backgrounds limit spatial-temporal methods, while weak target information and competing local or non-local structures challenge existing approaches. The paper proposes RIPT to combine complementary priors for more robust separation.

  • I. INTRODUCTION: Single-frame detection is important because rapid relative motion makes imaging backgrounds change quickly, degrading spatial-temporal methods.The task supports applications including early-warning, precision-guided weapon, missile-tracking, and maritime-surveillance systems.
  • I. INTRODUCTION: Small infrared targets provide little information because they lack texture and shape features, while brightness varies across target types, distances, and environments.This makes background characteristics and target-background relations especially important for detection.
  • I. INTRODUCTION: Existing methods mainly exploit either local pixel correlation or non-local patch self-correlation, leaving their complementary strengths insufficiently combined.Local methods use nearby-pixel consistency, whereas non-local methods model background patches through low-rank recovery.
  • I. INTRODUCTION: Local-prior methods can miss dim targets through rare-structure effects, while non-local-prior methods retain salient edge residuals that can obscure dim targets.The paper identifies these as distinct failure modes of relying on only one prior.
  • I. INTRODUCTION: RIPT combines non-local redundancy and local structure through an infrared patch-tensor model, reweighted element-wise weighting, and sparsity enhancement.The framework is optimized for target-background separation and is reported to produce robust, clear target images across scenarios.

II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL

The IPT model stacks image patches into a spatially structured tensor and separates low-rank background, sparse targets, and noise. It then motivates entry-wise reweighting because a global sparsity trade-off cannot simultaneously preserve dim targets and suppress strong edges.

  • II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL: The IPT model transfers the image decomposition into patch space by stacking sliding-window patches into a 3D tensor.F, B, T, and N denote input, background, target, and noise patch-tensors; I and J are patch dimensions, and P is the patch count.
  • II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL: Non-local background patches can be equivalent across image regions, supporting low-rank structure in all three unfolding modes of the background patch-tensor.The rapidly decreasing singular values shown for modes 1–3 motivate using tensor structure beyond a single patch-image unfolding.
  • II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL: The target patch-tensor is modeled as extremely sparse because small targets occupy only a few pixels in the image.The parameter k represents a small quantity determined by target number and size.
  • II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL: Higher-order robust tensor recovery replaces intractable tensor rank and l0 sparsity with Tucker-rank nuclear-norm and l1 surrogates.The resulting convex formulation sums nuclear norms over the mode-i unfoldings and includes a bounded additive-noise assumption.
  • II. NON-LOCAL CORRELATION DRIVEN INFRARED PATCH-TENSOR MODEL: A globally constant λ cannot both preserve dim targets and suppress strong cloud edges, motivating entry-wise weighting.Larger λ removes sparse non-target components but can over-shrink targets; smaller λ retains targets but also keeps edges.

III. INCORPORATING LOCAL STRUCTURE PRIOR

The model incorporates local structure information through a structure-tensor-based weight map, complementing the non-local correlation prior. This adaptive weighting suppresses strong edges while preserving dim targets.

  • III. INCORPORATING LOCAL STRUCTURE PRIOR: The local structure weight is constructed from the image structure tensor and interpreted as an edge-salience measure.The structure tensor estimates local geometry, including edge orientation.
  • III. INCORPORATING LOCAL STRUCTURE PRIOR: The eigenvalue difference λ1 − λ2 serves as a low-cost edge-awareness feature because it is larger at edges than in flat or corner regions.Flat regions have λ1 ≈ λ2 ≈ 0, edges have λ1 ≫ λ2 ≈ 0, and corners have λ1 ≥ λ2 ≫ 0.
  • III. INCORPORATING LOCAL STRUCTURE PRIOR: The structure-tensor feature maps are transformed into patch-tensors and used to construct the local structure weight.The resulting maps are illustrated as edge-awareness maps for the images in Fig. 5.
  • III. INCORPORATING LOCAL STRUCTURE PRIOR: The local weight yields a weighted IPT model in which strong edges are suppressed in the target image.The weighting is applied through the Hadamard product.

A. Reweighted infrared patch-tensor model

The reweighted IPT model enhances target sparsity and combines it with local-structure adaptation and a non-negativity prior. An additional sparsity-based stopping condition is motivated by the target tensor’s behavior during iterations.

  • A. Reweighted infrared patch-tensor model: The additional stopping condition is intended to reduce computing time because target sparsity stabilizes before full reconstruction convergence.The paper illustrates this reduction in Fig. 10 and Table IV.
  • A. Reweighted infrared patch-tensor model: The reweighted scheme assigns large weights to discourage non-zero target entries and small weights to preserve them.This scheme is adopted from reweighted ℓ1 minimization to enhance sparsity.
  • A. Reweighted infrared patch-tensor model: The algorithm can stop when the target tensor’s non-zero entry count stabilizes, in addition to satisfying the relative-error criterion.The condition is ∥T k+1∥0 = ∥T k∥0.
  • A. Reweighted infrared patch-tensor model: The target tensor is constrained to have non-negative entries because infrared small targets are brighter than their neighborhood environment.This prior is incorporated through an indicator-function constraint.
  • A. Reweighted infrared patch-tensor model: The local structure weight and sparsity-enhancing weight are combined into an adaptive weight for the reweighted IPT model.The resulting model is named the reweighted infrared patch-tensor model, or RIPT.

B. Solution of RIPT model

RIPT is solved as a reweighted robust tensor recovery problem using ADMM. The adaptive weights control soft-thresholding, balancing target preservation against edge suppression.

  • B. Solution of RIPT model: The proposed RIPT model is solved as a reweighted robust tensor recovery problem via the Alternating Direction Method of Multipliers.ADMM decomposes the augmented-Lagrangian minimization into background and target subproblems.
  • B. Solution of RIPT model: ADMM alternates updates of the background tensors, target tensor, and Lagrange multipliers.The paper presents separate updates for Bi, T, and Yi.
  • B. Solution of RIPT model: The subproblems are solved using two corresponding operators within the iterative optimization procedure.The operators are associated with the background and target updates.
  • B. Solution of RIPT model: The element-wise adaptive weight determines the soft-threshold and therefore controls the trade-off between the target and background patch-tensors.This enables simultaneous preservation of small targets and suppression of strong edges.

C. Detection Procedure

The detection procedure computes local structure features, constructs patch-tensors, decomposes them with RIPT, reconstructs target and background images, and segments the target using an adaptive threshold.

  • C. Detection Procedure: The procedure first computes a local structure feature map and constructs the original patch-tensor and local structure weight patch-tensor.These steps use the input infrared image and its local structure feature map.
  • C. Detection Procedure: Algorithm 1 decomposes the image patch-tensor into background and target patch-tensors using RIPT.The algorithm initializes the tensors and iteratively updates the model variables until convergence.
  • C. Detection Procedure: Each iteration updates the background tensors, target tensor, multipliers, adaptive weight, and penalty parameter before checking convergence.The convergence check includes the algorithm’s stopping conditions.
  • C. Detection Procedure: The reconstructed background and target images are obtained from their corresponding patch-tensors using the uniform average of estimators reprojection scheme.The target is then segmented as in Ref..
  • C. Detection Procedure: The segmentation threshold is adaptive and depends on the target image’s average and standard deviation together with empirically determined constants.The threshold uses f̄T, σ, k, and vmin.

V. EXPERIMENTAL VALIDATION

The experiments evaluate the proposed algorithm across varied infrared scenarios against ten state-of-the-art methods.

  • The proposed algorithm is evaluated using images from various scenarios.

A. Experimental setup

The experimental setup uses real infrared sequences spanning simple to highly cluttered scenes, compares twelve methods, and evaluates background suppression with four metrics.

  • Real infrared images cover flat backgrounds with salient targets, complex backgrounds with heavy clutters, and extremely dim targets.
  • The experiments focus on complex scenes containing very dim targets and strong clutters because they are the most challenging cases.
  • The proposed method is compared with twelve methods spanning filtering-based, HVS-based, and low-rank approaches.
  • Evaluation uses LSNRG, BSF, SCRG, and ROC curves to compare background-suppression performance.
  • LSNRG measures local signal-to-noise-ratio gain before and after background suppression.
  • BSF measures background suppression using the standard deviation of the neighborhood region.
  • SCRG is the ratio of signal-to-clutter ratios before and after processing, with higher LSNRG, BSF, and SCRG indicating better suppression.
  • The ROC curve represents the trade-off between true detections and false detections.

B. Validation of the proposed method

The validation examines robustness across scenes and noise, then isolates the contributions of the patch-tensor, local-structure, and sparsity-enhancement priors.

  • The experiments assess robustness to various scenes and noisy cases while examining each introduced prior individually.

1) Robustness to various scenes:

The proposed RIPT method separates targets across varied scenes, remains effective under moderate noise, and benefits mainly from patch-tensor and local-structure priors.

  • Robustness to various scenes: The separated target images show background clutters removed across the varied scenarios, indicating robustness to different scenes.
  • Robustness to noise: With noise standard deviation 10, the method enhances targets and suppresses clutters and noise; at 20, it detects some targets but fails when noise overwhelms them.
  • Robustness to noise: Noise impact depends on both noise intensity and the target’s original contrast, with weak residual contrast sufficient for detection in some cases.
  • Roles of components in the proposed model: Patch-tensor methods outperform IPI, demonstrating that mode-1 and mode-2 unfolding matrices contribute to detection performance.
  • Roles of components in the proposed model: Incorporating the local structure prior significantly improves detection probability, while sparsity enhancement reduces iteration count without improving final detection performance.
  • ADMM provides O(1/k) convergence for the proposed solver, while computational time differs among methods despite similar apparent complexities.
  • Parameters analysis: Patch sizes from 20–60 produce acceptable detection results, and size 30 performs best on Sequences 1–4.

2) Sliding step:

The model’s parameter studies show trade-offs between computational redundancy, clutter suppression, target preservation, and detection performance, while comparisons favor RIPT in complex scenes.

  • Sliding step: Sliding step 12 is selected because larger steps reduce computation but can weaken patch redundancy and detection quality.The algorithm remains robust to step-length variation, although step 8 produces a sharper ROC curve with relatively low overall detection probability.
  • Weight stretching parameter h: The weight-stretching parameter h balances clutter-edge suppression against preserving targets that move along cloud edges.Large h values can suppress dim targets mistaken for edges, whereas small h values retain more non-target components and increase false alarms.
  • Weighting parameter λ: The weighting parameter must avoid both excessive thresholds that suppress dim targets and insufficient weighting that raises false alarms.Large λ values keep false-alarm ratios low but reduce detection probabilities, while too-small L values produce higher false-alarm ratios.
  • Penalty factor μ: The penalty factor μ controls background low-rank shrinkage, trading preservation of background details against leakage of targets or non-target components.Smaller μ preserves more background details and may absorb targets, whereas larger μ leaves more non-target components in the target patch-tensor.
  • Quantitative evaluation: The reported local suppression indices show strong regional performance but do not necessarily establish globally effective background suppression.This qualification limits how LSNRG, BSF, and SCRG should be interpreted in the comparison.
  • Comparison with State-of-the-Arts: RIPT achieves the highest detection probability at the same false-alarm ratio and produces clearer target separation than the compared state-of-the-art methods.The comparison includes twelve tested methods across representative sequences; local and low-rank baselines retain more clutter or edges in difficult scenes.

VI. CONCLUSION

The paper concludes that RIPT combines non-local redundancy and local structure information through a reweighted infrared patch-tensor model for robust single-frame small-target detection.

  • VI. CONCLUSION: RIPT combines non-local redundant and local structure priors to suppress strong edges while preserving spatial correlations.A structure-tensor-based local weight and sparsity enhancement are incorporated into a reweighted robust tensor recovery problem solved via ADMM.
Loading 1703.09157v1…