The deployment of deep learning models in high-stakes applications such as autonomous driving is critically important, yet their black-box nature remains a fundamental barrier to trust and accountability. Existing explainability methods typically produce ambiguous, pixel-based heatmaps that capture correlation rather than establishing a causal link between high-level, human-interpretable concepts and model outputs. This paper introduces the Causal Concept Decomposer (CCD), a three-stage framework for concept-driven causal explanation of object detectors. CCD first employs semantic segmentation to isolate the target object, then applies Non-negative Matrix Factorization to discover constituent semantic parts, and finally uses Sobol sensitivity analysis to quantify the causal influence of each part on the detector’s decision. Evaluated on the MS COCO dataset, CCD produces explanations that are both visually coherent and quantitatively more faithful than existing approaches, achieving a Deletion AUC of 0.11 and an Insertion AUC of 0.91. By moving beyond correlational attribution towards principled causal analysis, this work represents an important step towards more trustworthy and reliable AI systems.

Abstract Large Multimodal Models (LMMs) have achieved remarkable performance across vision-language tasks, yet their robustness against adversarial attacks remains critically underexplored. While LMMs are vulnerable to visual encoder attacks, they exhibit surprising resilience due to encoder diversity—attacks optimized for CLIP fail to transfer to EVA-CLIP, especially when textual context is provided. We introduce the Adaptive Ensemble PGD (AE-PGD) attack, which simultaneously targets both encoders through three key innovations: (1) dynamic adversarial caption selection , combining gradient magnitude with global semantic displacement to identify the most attack-effective caption per model; (2) an adaptive weight controller , dynamically balancing each encoder’s contribution using real-time loss, gradient norm, and confidence metrics; and (3) an Expectation over Transforms (EoT) gradient update ensuring robustness against input-transformation defenses. Evaluated on COCO 2014 images, AE-PGD reduces accuracy from a 75.42% baseline to 0.0% across all three evaluation metrics—visual encoding, image-to-text recall, and LLM answer recall—achieving complete model collapse. Manifold analysis confirms that adversarial perturbations push image embeddings to antipodal regions of the joint embedding space, activating semantically opposite concept clusters and producing structured hallucinations. WordNet WUP similarity analysis reveals a 33.5 percentage point semantic drop across the test set. AE-PGD causes state-of-the-art LMMs (LLaVA, Qwen-VL, GPT-4V) to catastrophically misidentify a bullet train as a “helicopter crash,” with strong black-box transfer yielding a 65 percentage point recall collapse on unseen encoders. This work exposes critical vulnerabilities in current LMM architectures and underscores the urgent need for ensemble-aware defense mechanisms.

Sparse Coding (SC) and Convolution Sparse Coding (CSC) are two widely studied sparse methods in computer vision and signal processing. SC encodes the image patches independently, however fails to utilize the correlation among them. CSC adopts a convolution operator to connect the overlapping patches but in an inflexible manner. In this paper, a novel integrated framework for the two sparse models is proposed, wherein the local correlations among patches are controllable by manipulating a window function. Moreover, the inherent border effect of a convolution model is mitigated with a carefully designed weight function. It can be demonstrated that both SC and CSC are two distinct implementations of this framework. Consequently, our unified framework provides a balanced solution by addressing the strengths and limitations of both SC and CSC. Extensive experimental results are presented to demonstrate the superiority and effectiveness of the proposed method for image inpainting tasks.

In scenarios where identifying face information in the visible spectrum (VIS) is challenging due to poor lighting conditions, the use of near-infrared (NIR) and thermal (TH) cameras can provide viable alternatives. However, the unique data distribution of images captured by these cameras compared to VIS images presents challenges in matching face identities. To address these challenges, we propose a novel image transformation framework. The framework includes feature extraction from the input image, followed by a transformation network that generates target domain images with perceptual fidelity. Additionally, a reconstruction network preserves original information by reconstructing the original domain image from the extracted features. By considering the correlation between features from both domains, our framework utilizes paired data obtained from the same individual. We apply this framework to two well-established image-to-image transformation models, pix2pix and CycleGAN, known as CRC-pix2pix and CRC-CycleGAN respectively. The versatility of our approach allows extension to other models based on pix2pix or CycleGAN architectures. Our models generate high-quality images while preserving the identity information of the original face. Performance evaluation on TFW and BUAA NIR-VIS datasets demonstrates the superiority of our models in terms of generated image face matching and evaluation metrics such as SSIM, MSE, PSNR, and LPIPS. Moreover, we introduce the CQUPT-VIS-TH dataset, which enriches the paired dataset with thermal-visual face data capturing various angles and expressions.

Eye Center Localization (ECL) is one of the most crucial technologies for various computer vision applications, such as eye gazing estimation and eye-tracking. Current conventional implementations consist of two phases, including locating the approximate eye regions and finding the eye center position by extracting the semantic features around the corresponding eye region. However, the combination pipeline results in the ECL accuracy being influenced by not only the environmental factors, such as the variability of photographing angles, illuminations, and the probable occlusions by eyelids or glasses, but also the quality of preceding procedures. Inspired by the ensemble mechanism in machine learning, we formulate ECL problem as a process of end-to-end voting, and the core is to select a set of local descriptors which can capture efficient independent information to vote for eye centers. With the help of deep convolutional neural networks, we are able to determine semantic descriptors around the eye regions. Each descriptor proposes a vote pointing to the corresponding eye center, and all the votes indicate the eye centers finally. The experimental results on the public databases, BioID and GI4E, show that our method achieves 80.3% and 95.2% accuracy, respectively, which outperforms the existing state-of-the-art methods, and the results based on our customized challenging database verify the robustness of our method.

Current state-of-the-art detectors achieved impressive performance in detection accuracy with the use of deep learning. However, most of such detectors cannot detect objects in real time due to heavy computational cost, which limits their wide application. Although some one-stage detectors are designed to accelerate the detection speed, it is still not satisfied for task in high-resolution remote sensing images. To address this problem, a lightweight one-stage approach based on YOLOv3 is proposed in this paper, which is named Squeeze-and-Excitation YOLOv3 (SE-YOLOv3). The proposed algorithm maintains high efficiency and effectiveness simultaneously. With an aim to reduce the number of parameters and increase the ability of feature description, two customized modules, lightweight feature extraction and attention-aware feature augmentation, are embedded by utilizing global information and suppressing redundancy features, respectively. To meet the scale invariance, a spatial pyramid pooling method is used to aggregate local features. The evaluation experiments on two remote sensing image data sets, DOTA and NWPU VHR-10, reveal that the proposed approach achieves more competitive detection effect with less computational consumption.

In linear representation-based image classification, an unlabeled sample is represented by the entire training set. To obtain a stable and discriminative solution, regularization on the vector of representation coefficients is necessary. For example, the representation in sparse representation-based classification (SRC) uses L1 norm penalty as regularization, which is equal to lasso. However, lasso overemphasizes the role of sparseness while ignoring the inherent structure among samples belonging to a same class. Many recent developed representation classifications have adopted lasso-type regressions to improve the performance. In this paper, we propose the adaptive class preserving representation for classification (ACPRC). Our method is related to group lasso based classification but different in two key points: When training samples in a class are uncorrelated, ACPRC turns into SRC, when samples in a class are highly correlated, it obtains similar result as group lasso. The superiority of ACPRC over other state-of-the-art regularization techniques including lasso, group lasso, sparse group lasso, etc. are evaluated by extensive experiments.

Error detection is an important approach to improve the robustness of face recognition method. However, it is hard to directly detect the invalid pixels in a facial image. The authors decompose the hard problem into many simpler sub‐problems in this study. That is, the error detection process of pixels is divided into multiple phases and a portion of invalid pixels are detected in each phase. The goal is to decrease the ratio of invalid pixels to the whole pixels in a testing image, which progressively improves the final recognition accuracy. The performance that their method deals with occlusion and corruption problems is evaluated on different databases. In addition, the comparison with other state‐of‐the‐art studies shows that the proposed method achieves the best results in face occlusion and disguise issues.

How to identify a set of genes that are relevant to a key biological process is an important issue in current molecular biology. In this paper, we propose a novel method to discover differentially expressed genes based on robust principal component analysis (RPCA). In our method, we treat the differentially and non-differentially expressed genes as perturbation signals S and low-rank matrix A, respectively. Perturbation signals S can be recovered from the gene expression data by using RPCA. To discover the differentially expressed genes associated with special biological progresses or functions, the scheme is given as follows. Firstly, the matrix D of expression data is decomposed into two adding matrices A and S by using RPCA. Secondly, the differentially expressed genes are identified based on matrix S. Finally, the differentially expressed genes are evaluated by the tools based on Gene Ontology. A larger number of experiments on hypothetical and real gene expression data are also provided and the experimental results show that our method is efficient and effective.

Constrained independent component analysis (cICA) is a general framework to incorporate a priori information from problem into the negentropy contrast function as constrained terms to form an augmented Lagrangian function. In this letter, a new improved algorithm for cICA is presented through the investigation of the inequality constraints, in which different closeness measurements are compared. The utility of our proposed algorithm is demonstrated by the experiments with synthetic data and electroencephalogram (EEG) data.

This paper presents a comparison of three kinds of principal component neural networks, which are used for image compression. Principal component analysis (PCA), which is a statistical processing technique, is used in many engineering and scientific fields. Application of computing principal components includes data compression, pattern recognition and signal processing, etc. In this paper we use it for image compression.