In a real-world infrared (IR) imaging system, effectively learning a consistent stripe noise removal model is essential. Most existing destriping methods cannot precisely reconstruct images due to cross-level semantic gaps and insufficient characterization of the global column features. To tackle this problem, we propose a novel IR image destriping method, called asymmetric sampling correction network (ASCNet), that can effectively capture global column relationships and embed them into a U-shaped framework, providing comprehensive discriminative representation and seamless semantic connectivity. Our ASCNet consists of three core elements: residual Haar discrete wavelet transform (RHDWT), pixel shuffle (PS), and column nonuniformity correction module (CNCM). Specifically, RHDWT is a novel downsampler that employs double-branch modeling to effectively integrate stripe-directional prior knowledge and data-driven semantic interaction to enrich the feature representation. Observing the semantic patterns crosstalk of stripe noise, PS is introduced as an upsampler to prevent excessive a priori decoding and performing semantic-bias-free image reconstruction. After each sampling, CNCM captures the column relationships in long-range dependencies. By incorporating column, spatial, and self-dependence information, CNCM well establishes a global context to distinguish stripes from the scene’s vertical structures. Extensive experiments on synthetic data, real data, and IR small target detection (IRSTD) tasks demonstrate that the proposed method outperforms state-of-the-art single-image destriping methods both visually and quantitatively. The code is available athttps://github.com/xdFai/ASCNet.

Deep learning (DL) networks have recently achieved excellent performance on image compressed sensing. However, most existing methods rely on burdened and complex network structures, resulting in significant computational and storage requirements that defeat the purpose of compressed sensing. This severely hinders their applicability in real-world resource-limited devices. In this paper, a lightweight cycle network driven by physical and deep priors for image compressed sensing is proposed which integrates the learning of the sensing matrix and compressive image reconstruction. Specifically, the regularization terms and a likelihood term derived from the physical observation model are learned in an end-to-end cycle network, simultaneously estimating the reconstructed image and sensing matrix in the image and feature domains. Moreover, a dual-domain fusion reconstruction module is proposed. It creates simulated measurement residuals for enhancing reconstruction in the compressed domain, which leads to high reconstruction performance and reduces computational load by bonding together the compressed image domains in the cyclic network. Extensive experiments demonstrate that our model delivers superior performance and alleviates model complexity, which is of great importance in low-budget applications.

Hyperspectral images (HSIs) often suffer from degradation caused by mixed noise, leading to a decline in the performance of subsequent advanced applications. To eliminate noise and improve image quality, transformer-based approaches have been successfully employed. Nevertheless, these strategies often involve large-scale modeling and tedious layer normalization, which causes inefficiencies during the denoising process. Additionally, the neglect of local spectral correlations in HSIs damages the physical properties in recovery, resulting in poor generalization and inefficient denoising performance. To address these problems, we propose an efficient spatial–spectral oriented triple attention network, dubbed S2OTAN, for HSI denoising. Specifically, to fully exploit the physical properties of HSIs, we impose spatial and spectral multiscale hybrid attention in the single-transformer block side-by-side to fuse spatial–spectral information in a parallel manner. For spatial feature extraction, we introduce hybrid spatial attention by constructing attention maps for pixels within and across windows to exploit the local and global similarity in spatial and improve computational efficiency. For spectral feature exploration, we utilize spectral partitioning operations to enhance the adjacent spectral dependences of HSIs and capture contextual information related to correlations. Consequently, our method exhibits a robust feature representation capability for removing mixed noise in HSIs. Extensive experiments on synthetic and real-world noisy scenarios demonstrate that the proposed approach outperforms other state-of-the-art approaches among quantitative metrics and visual effects. For the sake of reproducibility, the code is available at:https://github.com/Zilong-Xiao/S2OTAN.

Improving hyperspectral image (HSI) quality is crucial in subsequent applications. Current transformer-based methods effectively remove mixed noise from the original HSIs. However, there is limited research on targeted modeling of the spatial locality and edge properties of HSI. In addition, the original transformer computes global query-key pairs indifferently, resulting in equal weights for dissimilarity features, noise, and essential information, thus interfering with the restoration of clean images. To address these issues, this study proposes an effective HSI denoising network called multiscale-sparse spatial-spectral transformer (MS3T) to achieve end-to-end mixed noise removal. Specifically, we reconstruct the attention module in the original transformer by adopting a dual-stream approach to selectively explore the 3-D information of HSI from both spatial and spectral perspectives. In the spatial domain, we construct a multiscale context-capturing module based on the local and non-local similarity properties of HSI to establish remote connections from local to global. In the spectral domain, we develop a top-k selection operator to calculate the similarity scores of query-key pairs and select important semantic information for efficient feature aggregation. Both of the above modules alleviate the computational complexity issue of the original transformer to different extents, and improve the mixed noise removal performance of the whole network. To validate the effectiveness and efficiency of MS3T, we conduct synthetic and real experiments on multiple datasets, and the final results demonstrate that our method outperforms the state-of-the-art methods in both metric evaluation and visual effects; the reproducible code is available athttps://github.com/Zilong-Xiao/MS3T.

Deep learning-based single-image infrared (IR) destriping has made significant advances. However, these methods are typically evaluated using “synthetic” images with specific stripe noise, making it unclear how well they handle “real” IR images. In fact, a clear and fair benchmarking of the existing destriping methods on real images, especially for the downstream IR small target detection (IRSTD) task, is currently an open gap. To tackle this problem, we introduce a novel benchmark, called IRSTD-oriented image destriping (IRSTDID-800), which thoroughly showcases the real distribution of IR small targets under stripe noise perturbation for the first time. Concretely, it consists of two subsets. (1) IRSTDID-SKY: composed of 500 real-world images afflicted with stripe noise, including unmanned aerial vehicles (UAVs) of various shapes, sizes, and contrasts. Moreover, these images are annotated with precise pixel levels for objective evaluation of IRSTD. (2) IRSTDID-GND: comprising 300 real-world images featuring common daily life objects, providing a richer scene under stripe noise. Based on the proposed IRSTDID-800, we comprehensively assess the performance of ten state-of-the-art (SOTA) destriping methods across eleven metrics, including full-reference, no-reference, and task-driven metrics with six advanced IRSTD methods. Furthermore, inspired by the correlation between image quality assessment and IRSTD, we proposed a task-oriented destriping optimization strategy. A loss function is introduced for IR image destriping, leveraging the structural properties of noise as a penalty term to strengthen image destriping and IRSTD. Overall, our analysis reveals interesting observations to guide future research in destriping and IRSTD tasks. Our dataset is available athttps://github.com/xdFai/IRSTDID-800.

Histological images are often tainted with two or more stains to reveal their underlying structures and conditions. Blind Color Deconvolution (BCD) techniques separate colors (stains) and structural information (concentrations), which is useful for the processing, data augmentation, and classification of such images. Classical BCD methods rely on a complicated optimization procedure that has to be carried out on each image independently, i.e., they are not amortized methods. In contrast, once they have been trained, deep neural networks can be used in a fast, amortized manner on unseen inputs. Unfortunately, the lack of large databases of ground truth color and concentrations has limited the development of deep models for BCD. In this work, we propose a deep variational Bayesian BCD neural network (BCD-Net) for stain separation and concentration estimation. BCD-Net is trained by maximizing the evidence lower bound of the observed images, which does not require the use of ground truth examples of stains and concentrations. Results obtained using two multicenter databases (Camelyon-17 and a stain separation benchmark) demonstrate the effectiveness of BCD-Net in the stain separation tasks, while drastically reducing the computation time compared to classical non-amortized methods.

In order to effectively utilize spectral and object spatial information to improve tracking performance, we design an Hyperspectral Video (HSV) tracker, namely DSP-Net, to integrate the various prior information. The gradient difference between spectral vectors is explored to develop a clustering technique. The approach generates a binary mask containing target spectral information and appearance clues. A feature cache is introduced to store historical information. Additionally, the channel shift operation is used on the timing to capture the trajectory clues of the target. With the help of the non-local mechanism, the trajectory clues, appearance clues and spectral information of the target are finally integrated, using the designed spectral-spatial joint perception module to enhance the expression of the target. Experimental results show that DSP-Net outperforms state-of-the-art HSV trackers on existing dataset.