Gastrointestinal (GI) disorders often require diagnostic endoscopy to detect structural abnormalities and guide treatment decisions. However, automating structured clinical information extraction from GI endoscopy remains challenging due to the complex visual patterns, image acquisition heterogeneity, and reliability constraints. This work proposes an efficient, robustness-aware visual question answering (VQA) framework for generating key components of endoscopy reports. With images and textual questions as input, five state-of-the-art VLMs are fine-tuned on the Kvasir-VQA dataset using parameter-efficient fine-tuning (PEFT) techniques, including LoRA and QLoRA, for memory and computational efficiency, and PaliGemma 2 achieved the best overall VQA performance. However, robustness evaluation of the two best-performing VLMs revealed significant susceptibility to image, text, and gradient perturbations, notably under the projected gradient descent (PGD) attacks; PaliGemma 2 results in an attack success rate (ASR) of 60.4%. As a mitigation strategy, gradient-based adversarial training contributed to the best overall performance, achieving the highest clean-set VQA accuracy of 88.3% using PaliGemma 2, while the ASR remained below 15% across evaluated perturbations. Therefore, the proposed approach can extract key diagnostic answers from GI endoscopy, mitigating resource constraints and attack vulnerability, providing a foundation for an efficient and reliable clinician-assistive AI system for GI diagnostics through structured information extraction.
The accuracy of surrogate models plays a crucial role in solving computationally expensive multi-/many-objective optimization problems. In addition, the design of model-based optimization and the infill criterion critically affects the quality of solutions under a limited computational budget. In this paper, we propose an expensive multi-/many-objective evolutionary algorithm that evaluates only a subset of objective functions. In the proposed approach, a separate surrogate model is trained for each objective using its own dataset to improve prediction accuracy. A new environmental selection strategy, based on an exponential formulation of convergence and diversity, is employed to generate a set of candidate Pareto-optimal solutions. These candidates are then non-dominated sorted according to their distances to the optimal solution of each objective in the decision space, and the non-dominated solution on the distances with the maximum crowding distance in the objective space is selected for real evaluation on a subset of objectives. Experimental results on 3-, 5-, and 10-objective DTLZ and WFG benchmark suites show that the proposed algorithm outperforms six state-of-the-art surrogate-assisted optimization algorithms. Furthermore, it achieves superior performance on two practical engineering applications, confirming its effectiveness in real-world scenarios.
Generative adversarial networks (GANs) have achieved remarkable progress in image generation tasks, yet issues such as mode collapse and limited sample diversity constrain their broader applicability. This paper proposes a novel multi-objective conditional GAN that integrates residual attention mechanisms and topological constraints to address these challenges. The model employs an evolutionary optimization strategy based on two objective functions to encourage high image fidelity and mode diversity: a quality fitness score based on discriminator output, and a mode diversity score measuring intra-class variability. To enhance conditional generation, a novel generator architecture is designed by incorporating a conditional residual block that embeds label information via learnable feature-wise modulation. In addition, self-attention and channel attention mechanisms are employed to capture global structure and highlight class-relevant features. The discriminator adopts a symmetric architecture with attention-enhanced residual blocks and spectral normalization to ensure semantic consistency and stable adversarial training. Furthermore, a topological loss function based on persistent homology is introduced to enforce structural consistency by regulating the number of connected components and loops in the generated samples. Experimental results on MNIST, CIFAR-10, and a molecular dataset demonstrate that the proposed model achieves superior image fidelity and diversity compared to existing methods, effectively mitigating mode collapse while preserving class-specific characteristics.
For violence detection on event streams in privacy-constrained settings, many methods increase model complexity to gain accuracy, which conflicts with edge efficiency and interpretability. This work revisits time-series anomaly detection and introduces a two-stage system named TEV (Trees for Event-stream Violence Detection). The event-stream analysis is formulated as a weight-allocation problem over a long sequence defined by sliding windows. In Stage 1, the event-camera data are processed by a pretrained feature extractor. A dual-branch architecture conducts initial weight allocation, which is further refined using a graph self-attention network with a concatenation strategy. Center loss and video-level label loss are employed to produce both binary and multi-class predictions. In Stage 2, a tree-structured allocator is built on top of Stage 1 outputs. Unlike standard decision trees, the split function fuses feature and topological information and introduces a pointer mechanism that allows split nodes to reuse information computed in earlier splits, enabling task-adaptive behavior. Consequently, at inference time, the system uses the frozen feature extractor, compact graph readout/global head, and lightweight tree branch rather than a heavy video backbone. Experiments on multiple benchmark datasets show that the proposed TEV achieves higher performance than prior methods while remaining lightweight. Beyond video-level recognition, the evaluation includes frame-/segment-level AUC, AP, false-alarm analysis, component ablations, Top-$K$ and fusion-weight sensitivity, and on-device latency, memory, and power measurements on an NVIDIA Jetson Xavier AGX. These results show that the graph-to-tree migration improves localization, interpretability, and edge deployability without relying on a heavy video backbone.
Mars hyperspectral imaging enables mineralogical mapping by exploiting diagnostic absorption characteristics in high-dimensional reflectance spectra. However, compared with Earth HSI classification, which often benefits from richer land-cover annotations and stronger spatial or material semantics, few-shot Martian mineral classification is more constrained by scarce expert labels, high inter-class spectral similarity, and weak local absorption differences among minerals. Existing hyperspectral masked autoencoders mainly learn generic spectral–spatial continuity, but uniform reconstruction may underemphasize mineral-discriminative absorption structures in CRISM spectra. To address this issue, we propose an absorption-aware masked autoencoder for few-shot Mars hyperspectral classification. The framework preserves raw spectra through a raw reconstruction branch, emphasizes local absorption residuals through a continuum-guided branch, and introduces a derivative-aware constraint to retain spectral slope and absorption-edge variations. All reconstruction targets are derived from unlabeled Mars hyperspectral observations, so the pre-training remains self-supervised without using mineral labels. The learned encoder is then fine-tuned with only a few labeled pixels per class. Experiments on three HyMars scenes show that the proposed method improves accuracy and stability over supervised CNN/Transformer baselines, Mars-oriented methods, and generic hyperspectral masked modeling approaches, demonstrating the effectiveness of absorption-aware self-supervised pre-training for highly similar Martian mineral classes.
