Existing infrared physical adversarial attack methods struggle to balance strong attack performance and rapid deployment, with fixed iteration steps causing optimization redundancy. We propose an infrared adversarial patch optimization method based on a Gaussian heat diffusion model. By constructing aggregation regularization derived from Fourier's heat conduction law, we precisely guide digital-domain adversarial perturbations to form continuous aggregated shapes, improving physical realizability. We propose the number of connected regions as a compactness metric and, building upon it, design a dual-threshold early-stopping mechanism that further enhances optimization efficiency. Experiments on multiple infrared datasets demonstrate that our method outperforms mainstream approaches in both attack efficacy and optimization efficiency, and exhibits strong cross-model performance across detectors. Remarkably, physical experiments achieve the highest average attack success rate of 91.7% using only a single patch.

Low-resolution fine-grained image classification has recently made significant progress, largely thanks to the superresolution techniques and knowledge distillation methods. However, these approaches lead to an exponential increase in the number of parameters and computational complexity of models. In order to solve this problem, in this letter, we propose a Vision Mamba Distillation (ViMD) approach to enhance the effectiveness and efficiency of low-resolution fine-grained image classification. Concretely, a lightweight super-resolution vision Mamba classification network (SRVM-Net) is proposed to improve its capability for extracting visual features by redesigning the classification sub-network with Mamba modeling. Moreover, we design a novel multi-level Mamba knowledge distillation loss boosting the performance, which can transfer prior knowledge obtained from a High-resolution Vision Mamba classification Network (HRVMNet) as a teacher into the proposed SRVM-Net as a student. Extensive experiments on seven public fine-grained classification datasets related to benchmarks confirm our ViMD achieves a new state-of-the-art performance. While having higher accuracy, ViMD outperforms similar methods with fewer parameters and FLOPs, which is more suitable for embedded device applications. Code is available at Github.

Abstract Few‐shot learning (FSL) attempts to learn and optimise the model from a few examples on image classification, which is still threatened by data scarcity. To generate more data as supplements, data augmentation is considered as a powerful and popular technique to enhance the robustness of few‐shot models. However, there are still some weaknesses in applying augmentation methods. For example, all augmented samples have similar semantic information with respect to different augmented transformations, which makes these traditional augmentation methods incapable of learning the property being varied. To address this challenge, we introduce multi‐task learning to learn a primary few‐shot classification task and an auxiliary self‐supervised task, simultaneously. The self‐supervised task can learn transformation property as auxiliary self‐supervision signals to improve the performance of the primary few‐shot classification task. Additionally, we propose a simple, flexible, and effective mechanism for decision fusion to further improve the reliability of the classifier, named model‐agnostic ensemble inference (MAEI). Specifically, the MAEI mechanism can eliminate the influence of outliers for FSL using non‐maximum suppression. Extensive experiment results demonstrate that our method can outperform other state‐of‐the‐art methods by large margins.

Transformer-based architecture network has shown excellent performance in the coarse-grained image classification. However, it remains a challenge for the fine-grained image classification task, which needs more significant regional information. As one of the attention mechanisms, transformer pays attention to the most significant region while neglecting other sub-significant regions. To use more regional information, in this letter, we propose a complemental attention multi-feature fusion network (CAMF), which extracts multiple attention features to obtain more effective features. In CAMF, we propose two novel modules: (i) a complemental attention module (CAM) that extracts the most salient attention feature and the complemental attention feature. (ii) a multi-feature fusion module (MFM) that uses different branches to extract multiple regional discriminative features. Furthermore, a new feature similarity loss is proposed to measure the diversity of inter-class features. Experiments were conducted on four public fine-grained classification datasets. Our CAMF achieves 91.2%, 92.8%, 93.3%, 95.3% on CUB-200-2011, Stanford Dogs, FGVC-Aircraft, and Stanford Cars. The ablation study verified that CAM and MFM can focus on more local discriminative regions and improve fine-grained classification performance.

Background subtraction is a classic video processing task pervading in numerous visual applications such as video surveillance and traffic monitoring. Given the diversity and variability of real application scenes, an ideal background subtraction model should be robust to various scenarios. Even though deep-learning approaches have demonstrated unprecedented improvements, they often fail to generalize to unseen scenarios, thereby less suitable for extensive deployment. In this work, we propose to tackle cross-scene background subtraction via a two-phase framework that includes meta-knowledge learning and domain adaptation. Specifically, as we observe that meta-knowledge (i.e., scene-independent common knowledge) is the cornerstone for generalizing to unseen scenes, we draw on traditional frame differencing algorithms and design a deep difference network (DDN) to encode meta-knowledge especially temporal change knowledge from various cross-scene data (source domain) without intermittent foreground motion pattern. In addition, we explore a self-training domain adaptation strategy based on iterative evolution. With iteratively updated pseudo-labels, the DDN is continuously fine-tuned and evolves progressively toward unseen scenes (target domain) in an unsupervised fashion. Our framework could be easily deployed on unseen scenes without relying on their annotations. As evidenced by our experiments on the CDnet2014 dataset, it brings a significant improvement to background subtraction. Our method has a favorable processing speed (70 fps) and outperforms the best unsupervised algorithm and top supervised algorithm designed for unseen scenes by 9% and 3%, respectively.

Recent works have demonstrated that image descriptors produced by convolutional feature maps provide state-of-the-art performance for image retrieval and classification problems. However, features from a single convolutional layer are not robust enough for shape deformation, scale variation, and heavy occlusion. In this letter, we present a simple and straightforward approach for extracting multiscale (MS) regional maximum activation of convolutions features from different layers of the convolutional neural network. And we also propose aggregating MS features into a single vector by a parameter-free hedge method for image retrieval. Extensive experimental results on three challenging benchmark datasets indicate that the proposed method achieved outstanding performance against state-of-the-art methods.

Visual object tracking is a challenging task due to object appearance changes caused by shape deformation, heavy occlusion, background clutters, illumination variation, and camera motion. In this letter, we propose a novel robust algorithm which decomposes the task of tracking into translation and scale estimation. We estimate the translation by using five correlation filters with hierarchical convolutional features which produced multilevel correlation response maps to collaboratively infer the target location. We also calculate the scale variation by another correlation filter with histogram of oriented gradient features at the same time. Extensive experimental results on a large-scale 50 challenging benchmark dataset show that the proposed algorithm achieved outstanding performance against state-of-the-art methods.

Robust scale calculation is a challenging problem in visual tracking. Most existing trackers fail to handle large scale variations in complex videos. To address this issue, we propose a robust and efficient scale calculation method in tracking-by-detection framework, which divides the target into four patches and computes the scale factor by finding the maximum response position of each patch via color attributes kernelized correlation filter. In particular, we employ the weighting coefficients to remove the abnormal matching points and transform the desired training output of the conventional classifier to solve the location ambiguity problem. Experiments are performed on several challenging color sequences with scale variations in the recent benchmark evaluation. And the results show that our method outperforms state-of-the-art tracking methods while operating in real-time.