Active learning mitigates high annotation costs by selecting informative samples for annotation, yet most existing approaches rely on deterministic models that fail to effectively decouple epistemic and aleatoric uncertainties. To address this, a novel full-process probability-driven active learning (FPAL) framework is proposed, pioneering joint uncertainty and diversity queries in homogeneous probabilistic paradigm through a global guidance paradigm based on complete probability distributions with multiple components. Specifically, the Mixture Density Network is introduced for the first time in pathological diagnosis and integrated with Monte Carlo sampling for comprehensive uncertainty assessment. Meanwhile, the Jensen-Shannon divergence guided by complete probability distributions is designed for diversity query, enabling the FPAL not only to explore both under-represented feature regions and model-uncertain cognitive spaces. Extensive in-domain and cross-domain experiments on two public pathological datasets demonstrate that FPAL outperforms mainstream baselines, achieving higher accuracy with significantly fewer labeled samples. Code is available at: https://github.com/Norazm/FPAL-main.
The segmentation of lesions in dermoscopic images plays a crucial role in intelligent diagnosis. For issues such as varying sizes of skin lesion, blurred edges, complex boundaries, and low color contrast between the lesion and healthy skin, a model LKM-SAM was proposed for skin lesions segmentation. Firstly, fine-tuning the Segment Anything Model through Low-Rank Adaptation (LoRA), which involves freezing the parameters in the pre-trained model and using the trainable low-rank bypass matrices as the projection layer of Transformer block, enhances model performance in segmentation tasks while reducing training resources. Secondly, the convolution layers in the Neck layer of the image encoder is replaced by KAN convolutions, which extracts multi-scale global context information and compensates for the loss of local details. Finally, the mask decoder is replaced by Mask2fomer segmentation head. Mask2former can dynamically generate masks, thereby enhancing the model’s ability to perceive objects with different scales and low color contrast. Experimental results demonstrate that LKM-SAM achieves outstanding performance, with a Dice Similarity Coefficient (DSC) of 0.873, 0.881, and 0.934, and an Intersection over Union (IoU) of 0.931, 0.936, and 0.966 on the ISIC2017, ISIC2018 and PH2 datasets, respectively. Compared with current mainstream segmentation models, the IoU increased by 0.270, 0.276, and 0.192, respectively, verifying the high efficiency and strong generalization ability of the proposed model.
Generalized few-shot object detection (GFSOD) aims to detect base and novel classes simultaneously under limited annotations, playing an important role in real-world applications like autonomous driving and intelligent surveillance with emerging novel objects. Nevertheless, existing detectors suffer from two difficulties: unstable novel-class features without stable global semantic references, and semantic confusion between visually similar base-novel pairs that blurs decision boundaries. This paper presents CCD-GFSOD, a two-stage framework consisting of the Centroid-based Feature Calibration Module (CFCM) and Dynamic Discriminative Constraint Module (DDCM). CFCM constructs learnable global semantic centroids from base-class data using cross-batch exponential moving average to calibrate novel-class features and suppress intra-class dispersion during finetuning. DDCM imposes adaptive logit margin constraints on the most confusing categories identified by feature-prototype similarity, so as to sharpen inter-class boundaries. Experiments on PASCAL VOC and MS COCO show 54.1 https://github.com/hfy650/CCD-GFSOD .
The rapid expansion of the poultry industry requires automated disease screening systems to support poultry health management and food security. However, high-capacity CNN and Transformer models remain too computationally intensive for real-time deployment on edge devices. To address this deployment bottleneck, we propose a cross-architecture knowledge distillation framework, termed S2M-KD, for efficient poultry disease screening from fecal images. The proposed framework has three key designs. First, S2M-KD transfers knowledge from a Swin Transformer teacher to a MobileNetV3 student, enabling the lightweight CNN model to exploit the teacher’s global representation capability. Second, a response-based distillation objective is introduced to transfer softened class-response knowledge from the Transformer teacher to the CNN student by jointly using ground-truth labels and soft-target supervision. Third, the resulting student model provides an efficient deployment-oriented solution for real-time monitoring of poultry disease on edge devices. Extensive experiments validate the effectiveness of the proposed method. Compared with the strongest Transformer baseline, MaxViT, S2M-KD improves accuracy by 0.25, 0.42, and 1.09 percentage points on the three datasets, respectively. Moreover, compared with the 27.5M-parameter Swin-Tiny teacher, the distilled 4.2M-parameter student reduces the parameter count by approximately 84.7 224× 224 , while preserving an accuracy of 98.75 https://github.com/zhaoxuanding/S2M-KD.git to facilitate future research.
Histopathology image analysis is a major application area for deep learning, and the LC25000 dataset has become a standard benchmark for lung and colon cancer classification. Reported accuracies on this dataset routinely exceed 99
