PaperlayerView exact revision on arXiv

Source-linked AI summary

SCDM-HTD: Spectral-Wise Conditional Diffusion Model for Hyperspectral Target Detection

Dunbin Shen, Wenfeng Kong, Jianjun Liu, Xiaorui Ma, Hongyu Wang

arXiv:2610.32768v1cs.CV

TL;DR

HTD methods often compare spectral features while neglecting distribution learning, which may limit robustness in practical settings. SCDM-HTD uses conditional diffusion to generate a detection map from background and target confidence scores, and experiments on four benchmark datasets report effectiveness with quantitative and qualitative improvements.

  • Problem

    HTD methods primarily compare features and may neglect distribution learning for background-target separation, potentially limiting practical robustness.

    Source: §I · p. 2

  • Method

    SCDM-HTD diffuses background and target confidence scores conditioned on pixel spectra, then uses a lightweight MLP denoiser and composite separation loss to generate a detection map.

    Source: §I · p. 2

  • Results

    Experiments on four benchmark datasets demonstrate SCDM-HTD’s effectiveness, with a balance of target saliency and background suppression and strong detection performance reported.

    Source: §I · p. 2, 1) Qualitative Comparison · p. 9, 2) Quantitative Comparison · p. 9

  • Takeaways & Limitations

    A deterministic reverse non-Markov chain skips sampling steps, reducing the number required while maintaining promising detection performance for practical applications.

    Source: §I · p. 2

Abstract

from arXiv · show

Hyperspectral target detection (HTD) aims to locate and recognize objects at a fine-grained scale, playing a pivotal role in remote sensing. Despite significant progress, current HTD methods primarily focus on feature comparison while neglecting distribution learning for background-target separation, which can limit detection robustness in practical applications. This paper addresses this issue by introducing a spectral-wise conditional diffusion model that reformulates HTD as a task of generating a detection map. Specifically, we treat background and target confidence scores as training samples and use the corresponding pixel spectra as guiding conditions to capture the distributional discrepancy between background and target through diffusion and denoising. To enhance practical applicability, we propose a lightweight conditional denoising network based on multilayer perceptron to map noisy confidence scores to their ideal counterparts and adopt a deterministic reverse non-Markov chain to skip sampling steps, thereby accelerating the generation process. Furthermore, we introduce a composite loss to improve background-target separation. This loss integrates constrained energy minimization loss to enhance target saliency and suppress the background by maximizing the target response while minimizing average output energy, alongside binary cross-entropy loss to classify background and target using pixel-wise labels. Experimental results on four benchmark datasets demonstrate the effectiveness of the proposed method, with improvements in both quantitative metrics and qualitative visual effects.

I. INTRODUCTION

HTD seeks fine-grained target identification, but existing approaches may be limited by assumptions, handcrafted representations, or insufficient distribution learning. SCDM-HTD instead models background and target confidence-score distributions to generate a detection map.

A. HTD Methods Review

  • HTD identifies targets at pixel or subpixel scale; existing statistical, representation-based, and deep-learning methods have distinct assumptions or data and modeling limitations.Feature comparison without explicit distribution learning may limit robustness in practical applications.Source: §I · pp. 1–2

I. INTRODUCTION

  • SCDM-HTD conditions diffusion on pixel spectra to model background and target confidence-score distributions and directly generate a detection map.Its lightweight MLP denoiser and deterministic reverse non-Markov chain skip sampling steps; a composite CEM and BCE loss guides background-target separation.Source: §I · p. 2

1) Statistical-Based Methods:

  • Statistical HTD methods use distributional assumptions or energy minimization, but assumption mismatches with complex, non-Gaussian practical data can limit performance.CEM suppresses background while enhancing targets by minimizing average output energy and maximizing target response.Source: 1) Statistical-Based Methods · pp. 2–3

2) Representation-Based Methods:

Representation-based HTD methods use matrix or tensor decomposition and regularization, while deep-learning methods address scarce labels through several training strategies. Prior diffusion work in HTD mainly serves as an auxiliary tool; the paper formulates detection as confidence-map generation.

2) Representation-Based Methods:

  • Representation-based HTD methods estimate components through matrix or tensor decomposition with regularization, but handcrafted constraints may misfit datasets and nonlinear capacity may be limited.Unified representations can support subpixel detection by separating target and background components.Source: 2) Representation-Based Methods · p. 3

3) Deep Learning-Based Methods:

  • Deep-learning HTD methods use transfer learning, sample augmentation, or unsupervised learning to address scarce labels, but may face domain drift, distribution mismatch, or limited robustness.The reviewed unsupervised methods often rely on spectral feature comparison while overlooking distribution learning.Source: 3) Deep Learning-Based Methods · pp. 3–4

B. Diffusion Models

  • Prior HTD diffusion methods primarily support background-target separation or simulate spectral variability, leaving a fully end-to-end diffusion framework for HTD unexplored.The proposed model instead directly generates the detection map, conditioned on the original HSI.Source: B. Diffusion Models · p. 4

A. Problem Formulation

  • Given an HSI and known target spectrum, HTD classifies each pixel as target or background; this paper reframes detection as generating a confidence map from distributional discrepancy.The conventional confidence score indicates how likely a testing pixel matches the target spectrum.Source: A. Problem Formulation · p. 4

B. Overall Framework

SCDM-HTD formulates hyperspectral target detection as generating confidence scores conditioned on pixel spectra, then refines them into a detection map. It constructs background and target confidence samples and trains the denoiser to separate them.

B. Overall Framework

  • SCDM-HTD generates and iteratively refines confidence scores conditioned on pixel spectra to produce the detection map.At inference, the model starts from Gaussian noise and uses the original HSI as conditional guidance.Source: B. Overall Framework · pp. 4–5
  • The separation loss combines CEM loss to enhance target responses and suppress background with BCE loss to classify background and target pixels.The training pipeline also diffuses confidence-score pairs and denoises them using corresponding pixel spectra as conditions.Source: B. Overall Framework · pp. 4–5

C. Confidence Score Construction

  • Original pixels serve as background samples, while target samples are generated by randomly mixing them with the target prior using γ selected from [0, 0.4].The mixing strategy is intended to simulate subpixel targets while maintaining diverse and balanced training samples.Source: C. Confidence Score Construction · p. 5
  • Cosine similarities between background or generated target pixels and the target prior define negative and positive confidence scores, respectively.The confidence scores are the samples subsequently used in diffusion.Source: C. Confidence Score Construction · pp. 5–6

D. Confidence Score Diffusion

The confidence-score diffusion model learns background and target score distributions by progressively adding noise and reconstructing scores conditioned on pixel spectra.

D. Confidence Score Diffusion

  • The model diffuses both positive target and negative background confidence scores, then reconstructs them conditioned on each pixel's spectral characteristics.The reverse process begins with Gaussian noise and progressively recovers the scores.Source: D. Confidence Score Diffusion · p. 6

1) Forward Diffussion Process:

  • Forward diffusion applies shared Markov chains that progressively add Gaussian noise to batches of background and target confidence scores.The background and target scores are denoted x0 and y0, with corresponding noisy sequences xt and yt.Source: 1) Forward Diffussion Process · p. 6
  • At a large timestep T, the noisy confidence scores xT and yT approach standard Gaussian noise, and the reparameterization trick permits closed-form sampling from their forward distributions.The text omits the contrastive part of y for notational simplicity in the derivations.Source: 1) Forward Diffussion Process · p. 6

2) Reverse Sampling Process:

The reverse process approximates intractable score distributions with a learned conditional denoiser, then iteratively samples from noise to produce the detection result.

2) Reverse Sampling Process:

  • CDNet approximates the intractable reverse distributions with learned conditional transitions.The reverse sampling process is parameterized by the denoising network fθ.Source: 2) Reverse Sampling Process · p. 6
  • The reverse transition approximates qx(xt−1|xt, x0, Bp) with pθx(xt−1|xt, Bp), using a parameterized mean and variance.The mean and variance are specified for the reverse transition.Source: 2) Reverse Sampling Process · p. 6
  • After training, CDNet estimates the initial confidence score from noisy scores, and inference iteratively samples from Gaussian noise guided by the original HSI.The reverse process uses σt = √(1 − αt) for simplicity and consistency with the forward process.Source: 2) Reverse Sampling Process · p. 6

E. Spectral Conditional Denoising

The section presents a lightweight conditional denoising network for confidence scores and an objective that combines diffusion reconstruction with background-target separation.

1) Conditional Denoising Network:

  • CDNet uses a fully connected residual architecture to map noisy confidence scores to clean scores, conditioned on pixel spectra and timestep embeddings.Its two residual blocks fuse nonlinear features with timestep and spectral-condition embeddings; a residual connection helps prevent gradient vanishing.Source: 1) Conditional Denoising Network · p. 7

2) Objective Loss Function:

  • The diffusion objective derives from the ELBO and trains reconstructed confidence scores to approximate their original values.The prior-matching term has no learnable parameters, so optimization focuses on posterior matching and reconstruction.Source: 2) Objective Loss Function · p. 7
  • Diffusion reconstruction alone does not sufficiently separate background from target, and inaccurate augmented scores leave denoising weakly supervised.The proposed background-target separation loss combines CEM and BCE losses to address these issues.Source: 2) Objective Loss Function · p. 7
  • The CEM loss enhances target response while suppressing background output energy, rather than fully recovering confidence scores.Its constraint sets the target response to 1, while its optimization term minimizes background output energy.Source: 2) Objective Loss Function · p. 7
  • The BCE loss assigns target and background confidence scores labels 1 and 0, respectively, for pixel-wise classification.The combined loss is L = LCEM + λLBCE, with λ balancing the two terms.Source: 2) Objective Loss Function · p. 7

F. Reverse Sampling Acceleration

The reverse sampler replaces DDPM’s many-step Markov chain with a non-Markovian process that can skip steps. It estimates initial confidence scores and produces a deterministic detection map in a few steps.

F. Reverse Sampling Acceleration

  • The reverse process is reformulated as a non-Markovian chain that skips steps to accelerate sampling.The DDPM Markov chain typically requires many iterations, which does not meet practical detection efficiency demands.Source: F. Reverse Sampling Acceleration · p. 8
  • The reverse sampler estimates initial confidence scores, and setting η = 0 removes added random noise to make sampling deterministic.The sampler uses a subsequence of length S from the full reverse-time sequence.Source: F. Reverse Sampling Acceleration · p. 8
  • The accelerated process transforms Gaussian noise into a deterministic detection map in a few steps, improving efficiency for practical applications.The method reports that acceleration significantly reduces the required number of steps while maintaining promising detection performance.Source: F. Reverse Sampling Acceleration · p. 8

A. Datasets

The evaluation uses four publicly available datasets; the supplied passages describe three of them and their target scenes.

1) MUUL Gulfport :

  • CASI-1500 experiments on data acquired over the Gulf Park Campus of the University of Southern Mississippi use a 30×30-pixel subimage, where two solid brown cloth panels comprising 10 pixels are targets.The original scene has 325 × 337 pixels, with 64 spectral bands retained after removing noisy bands 1–4 and 69–72.Source: 1) MUUL Gulfport · p. 8

2) HYDICE:

  • HYDICE experiments use an 80 × 100-pixel subimage containing several vehicles that occupy 19 pixels.The source scene has 307 × 307 pixels at 1 m resolution, with 162 bands retained after band removal.Source: 2) HYDICE · p. 8

3) Airport-4:

  • Airport-4 contains three airplane targets occupying 60 pixels in a 100 × 100-pixel scene with 191 spectral bands.The dataset is part of ABU and was captured by the AVRIS sensor at 3.4 m spatial resolution.Source: 3) Airport-4 · p. 8

4) Urban-2:

The section describes Urban-2, the evaluation setup, and qualitative comparisons. SCDM-HTD is reported to balance target saliency and background suppression across the evaluated scenes.

4) Urban-2:

  • Urban-2 is a 100 × 100-pixel ABU scene with 207 spectral bands, 17.2 m spatial resolution, and 155 target pixels.Its targets are regularly arranged man-made objects.Source: 4) Urban-2 · p. 8

B. Compared Methods and Evaluation Metrics

  • Evaluation uses visual detection maps, ROC analyses, background-target separability diagrams, three ROC-curve AUCs, AUCOA, and AUCSNPR.The derived ROC curves assess detection effectiveness, background suppression, and target detectability.Source: B. Compared Methods and Evaluation… · p. 8

C. Implementation Specification

  • The proposed method uses five sampling steps for all datasets and is trained for 200 epochs; competing methods use their published optimal hyperparameters.Training uses AdamW, batch size 256, and 500 timesteps.Source: C. Implementation Specification · pp. 8–9

1) Qualitative Comparison:

  • SCDM-HTD maintains a balance between target saliency and background suppression, while its separability diagrams show a larger background-target margin and smaller background box than competing methods.The comparison also reports that OSP, CEM, CSCR, and WDCCR have weak background suppression, whereas BLTSC, HTD-Net, TSTTD, and OS-VAE tend to miss targets.Source: Figure 4 · p. 9, 1) Qualitative Comparison · p. 9

2) Quantitative Comparison:

SCDM-HTD leads on AUC(Pf,Pd) and combines target enhancement with background suppression. Parameter analyses identify settings for the CEM and BCE losses, while ablations support using both losses together.

2) Quantitative Comparison:

  • SCDM-HTD achieves the highest AUC(Pf,Pd) and competitive performance on four other AUC measures, with low computational cost from its lightweight model and accelerated sampling.Table II also reports running times; OSP and CEM are fastest because of their simplified computations.Source: 2) Quantitative Comparison · p. 9, Table II · p. 11

1) Effect of Parameter

  • The CEM parameter µ balances background suppression and target enhancement; performance peaks at µ = 10 across all four datasets and then generally declines as µ increases.The authors set µ to 10 for all datasets.Source: 1) Effect of Parameter · p. 9
  • The BCE weight λ controls its contribution to the loss; performance is stable or improves slightly at first, then declines, with dataset settings of 1, 10^-3, 1, and 10^-4.The settings correspond, respectively, to MULL Gulfport, HYDICE, Airport-4, and Urban-2.Source: 1) Effect of Parameter · p. 9

1) Effectiveness of Background-Target Separation Loss:

2) Effectiveness of Reverse Sampling Acceleration:

DDIM-based reverse sampling cuts the sampling steps while detection performance remains stable or declines slightly. The conclusion describes the broader SCDM-HTD approach and proposes pretrained generalized diffusion models as future work.

2) Effectiveness of Reverse Sampling Acceleration:

  • DDIM-based reverse sampling reduces steps from 500 to 5 and substantially cuts time across four datasets, while detection performance remains stable or declines slightly.The authors describe this acceleration as improving efficiency and practicality for real-world applications.Source: 2) Effectiveness of Reverse Sampling… · pp. 11–12, Table V · p. 12

V. CONCLUSION

  • SCDM-HTD generates detection maps from confidence scores conditioned on pixel spectra, using an MLP denoiser, a composite CEM-and-BCE loss, and deterministic non-Markov sampling.The conclusion reports strong objective metrics, visual quality, and real-time requirements.Source: §V · p. 12
  • Future work will explore a generalized diffusion model pretrained on public datasets and fine-tuned on specific datasets to improve generalization and detection efficiency.Source: §V · p. 12

Related papers

This paper cites

Loading 2610.32768v1…