pattern recognition and computer vision

To address the limited robustness to noise, poor cross-domain generalization, and weak interpretability of existing learning algorithms, this paper proposes a unified adaptive dynamic network architecture with three-tiered collaboration among convolutional neural networks (CNNs), traditional machine learning (ML), and large language models (LLMs). Methodologically, we integrate CNNs with ML to construct a hybrid base model and—novelly—introduce an LLM as a high-level semantic guidance and decision verification module, augmented by a noise-robust training strategy. Key contributions include: (1) the first CNN-ML-LLM collaborative framework, significantly improving noise resilience and cross-domain generalization; (2) enhanced model interpretability and end-to-end multi-task adaptability; and (3) empirical validation on image recognition and text generation tasks, achieving high accuracy (average +3.2%) and strong generalization in real-world applications such as medical image classification and financial risk prediction.

High barriers to adopting pre-trained models and a lack of empirical guidance for strategy selection hinder practical deployment in few-shot image classification and object detection. Method: We systematically compare linear probing versus fine-tuning across ResNet, MobileNet, and EfficientNet, and propose an end-to-end TensorFlow framework integrating multi-scale feature-space visualization (PCA, t-SNE, UMAP) to unify analysis of representation evolution. Contribution/Results: Linear probing significantly outperforms fine-tuning under extreme data scarcity (≤100 samples per class) while accelerating training by 3–5×. The framework enables high-accuracy, rapid deployment (<1 hour for fine-tuning) on standard benchmarks (ImageNet-1K, CIFAR-100), balancing beginner-friendly usability with expert-level extensibility. It bridges the gap between theoretical representation analysis and real-world engineering practice.

This work addresses motion-induced artifacts in unsupervised 3D reconstruction of dynamic scenes caused by moving foreground objects. To tackle this, we propose a novel method that jointly models unsupervised 2D object decomposition and 3D geometric consistency. Our approach introduces FlowCapsules—a flow-guided network for unsupervised foreground segmentation—and a transient-object-mask-driven robust optimization kernel that detects and excludes dynamic objects under multi-view geometric consistency constraints. This kernel subsequently guides weighted bundle adjustment and NeRF training. Crucially, our method is the first to jointly learn object-centric representations and scene-level geometry without any annotations or controlled capture conditions. Experiments demonstrate significant improvements in both SfM and NeRF reconstruction accuracy on casually captured dynamic scenes, effectively eliminating motion artifacts. The framework advances explicit object-aware 3D understanding in open-world vision applications.

To address performance limitations in object detection and semantic segmentation under complex scenarios—including occlusion, small objects, and cross-domain generalization—this paper proposes a novel multimodal detection paradigm synergizing large language models (LLMs). Methodologically, it systematically integrates CNNs, YOLOv5/v8, and DETR architectures into an LLM-augmented inference framework, augmented by scalable data pipelines, model pruning, and quantization, and evaluated via a multi-dimensional metric system based on mAP and mIoU. Key contributions include: (1) bridging the gap between traditional feature engineering and end-to-end deep learning; (2) introducing a dynamic context enhancement mechanism tailored for challenging environments; and (3) achieving state-of-the-art accuracy-efficiency trade-offs on COCO and ADE20K. The fully open-sourced, reproducible framework significantly improves model generalizability and robustness across diverse real-world conditions.

This work addresses the lack of human-interpretable concepts in intermediate-layer representations of CNNs. We propose an unsupervised post-hoc method that optimizes an orthogonal rotation in feature space to extract disentangled, concept-level interpretable basis vectors from sparsely thresholded activation responses. Unlike supervised approaches relying on manually annotated concepts, ours is the first purely unsupervised paradigm for discovering highly interpretable bases. We further introduce an improved interpretability metric and a concept-alignment analysis framework, validating our method across multiple CNN architectures and datasets. Experiments demonstrate that the rotated intermediate representations significantly outperform supervised basis extraction methods in both conceptual diversity and interpretability. Our results reveal an inherent limitation of supervised paradigms—namely, their restricted coverage of conceptual breadth—and open a new direction for model interpretability research. (149 words)

This work proposes a training-free object detection method tailored for scenarios involving minor data variations where model training and annotation are impractical, such as GUI automation testing. By leveraging a segmentation foundation model—e.g., SAM—to generate image segments and integrating classical feature engineering for object classification, the approach rapidly adapts to new targets or interface changes without any training or labeled data. Evaluated on an in-vehicle navigation icon detection task, the method achieves performance comparable to learning-based detectors like YOLO, while entirely eliminating the need for model training. This significantly reduces deployment time and cost, demonstrating strong practical utility through its efficiency and adaptability.

Performance anti-patterns are prevalent in training and inference of computer vision (CV) models, yet existing approaches struggle to automatically and precisely localize problematic segments within long-duration execution traces. Method: We introduce the first benchmark dataset for CV performance anti-pattern detection—comprising 600+ PyTorch execution traces spanning diverse hardware platforms and CV tasks—and propose an iterative detection paradigm combining lightweight temporal modeling for coarse screening with large language models (LLMs) for fine-grained classification and diagnostic feedback, thereby overcoming LLM context and reasoning limitations. Our method integrates PyTorch Profiler analysis, cross-platform support (CUDA/ROCm), and a standardized annotation protocol. Contribution/Results: The framework achieves significantly higher detection accuracy than unsupervised clustering and rule-based statistical methods, demonstrates strong generalization across classification, detection, segmentation, and generation tasks, and supports end-to-end localization alongside actionable optimization recommendations—enabling the first evaluable, reproducible, benchmarked anti-pattern detection for CV systems.

This work proposes a self-supervised feature learning method specifically designed for object detection to address the heavy reliance on large-scale annotated data. By pretraining the feature extractor on unlabeled data and guiding the model to focus on semantically informative object regions, the approach significantly enhances the representational capacity of the detector under limited annotation budgets. Experimental results demonstrate that the proposed method outperforms conventional ImageNet-pretrained models across multiple object detection benchmarks, achieving not only improved detection accuracy but also greater robustness and reliability.

This work addresses the lack of theoretical foundations for clustering-based self-supervised/unsupervised learning methods by establishing, for the first time, a rigorous theoretical connection to classical statistical mixture models—particularly Gaussian Mixture Models (GMMs). We propose SiamMM, a unified end-to-end deep learning framework that jointly optimizes contrastive representation learning, cluster assignment, and mixture model parameter estimation. Empirically, SiamMM achieves state-of-the-art clustering performance across multiple benchmarks. Crucially, by leveraging mixture model confidence calibration, it reliably identifies label-noisy samples in training data—a capability absent in prior clustering approaches. This work provides the first statistically grounded, interpretable foundation for unsupervised clustering and introduces a novel paradigm for diagnosing dataset quality via clustering outcomes.

This work addresses the challenge of discovering reusable data analysis skills through autonomous exploration under fully unsupervised conditions to enhance agent performance. The authors propose DataCOPE, a novel framework that, for the first time, enables completely unsupervised discovery of such skills. DataCOPE integrates a data analysis agent, an unsupervised validator—comprising an adaptive checklist and an answer consistency verifier—and a skill manager. Through mechanisms including exploration trajectory generation, contrastive learning, and validation-signal-driven skill distillation, the framework adaptively distills skills suitable for both report-oriented and reasoning-based data analysis. Evaluated on the Deep Data Research and DABStep benchmarks, DataCOPE achieves task score improvements of 9.71% and 32.30%, respectively, significantly outperforming existing methods.