Logit Knowledge Distillation has gained substantial research interest in recent years due to its simplicity and lack of requirement for intermediate feature alignment; however, it suffers from limited interpretability in its decision-making process. To address this, we propose implicit Clustering Distillation (iCD): a simple and effective method that mines and transfers interpretable structural knowledge from logits, without requiring ground-truth labels or feature-space alignment. iCD leverages Gram matrices over decoupled local logit representations to enable student models to learn latent semantic structural patterns. Extensive experiments on benchmark datasets demonstrate the effectiveness of iCD across diverse teacher-student architectures, with particularly strong performance in fine-grained classification tasks---achieving a peak improvement of +5.08% over the baseline.

Detecting electron microscope images of cashmere and wool fibers is challenging; traditional manual methods struggle to distinguish between them. Cashmere and wool fibers are extremely similar in scale, fiber length, and fineness, coupled with the scarcity of relevant resources caused by the exorbitant cost of obtaining electron microscope images. These factors are intertwined, making it also difficult for the accuracy of deep learning detection methods to reach an ideal level. In response to the preceding issue, this paper presents a detector capable of learning discriminative features between fibers from electron microscope images and effectively completing the detection task with limited data. Initially, to address the challenge of capturing the complex features of cashmere and wool, this paper introduces Wool-Linear Deformable Convolution (W-LDConv) derived from LDConv. Furthermore, to overcome issues such as the inability of deformable large-kernel attention to adapt to fiber distribution and the excessive number of parameters, W-LDConv is used to take over the role of deformable convolution, forming the Linear Deformable Feature Capture Module (LDFC). Ultimately, the LDFC module is integrated with the efficient and stable YOLO11 model to propose the LDFC-YOLO object detector. Compared with the baseline model, the precision, recall, and [email protected] have improved by 12.9%, 18.6%, and 13.9%, respectively.