PotatoRelSeg: Intra-tuber Relative Representation Learning for Lightweight Potato Surface Defect Semantic Segmentation
Аннотация
Accurate potato surface defect segmentation is important for automated produce grading and post-harvest quality inspection, but remains challenging because defect regions are sparse and healthy-skin appearance varies substantially across individual tubers. This paper proposes PotatoRelSeg, a lightweight semantic segmentation framework based on intra-tuber relative representation learning. Instead of relying solely on absolute visual features or a shared class-level prototype, the framework estimates a sample-specific healthy-skin reference from coarse predictions and an image-derived foreground mask without requiring ground-truth annotations at inference. Pixel features are then represented jointly by their original descriptors, feature differences, cosine similarity, and Euclidean distance relative to the healthy reference and are refined through a compact depthwise-separable residual module. Healthy-reference alignment and local healthy-ring contrast further provide training supervision for the image-level reference and local defect–healthy relationships. Experiments on the predefined Diamant Potato Dataset split show that PotatoRelSeg achieves the strongest overall and defect-oriented segmentation performance among the evaluated methods. Sequential ablations further support the contribution of the proposed components, while efficiency analysis shows that the framework retains a compact model structure with a moderate computational overhead. These results suggest that intra-tuber reference conditioning provides a useful domain-specific inductive bias for potato surface defect segmentation.