Wuxi University model compares each tuber with its own skin to map defects
A Wuxi University researcher in China has described PotatoRelSeg, a lightweight image-segmentation model for potato surface defects, in Potato Research on 7 Nov 2026. Rather than comparing pixels with a fixed idea of healthy skin, it estimates a healthy-skin reference for each tuber, aiming to improve automated grading and post-harvest inspection.
Why it matters
Optical sorters on fry, chip and packing lines still struggle when skin color and texture vary between lots, which drives false rejects and missed defects. A per-tuber reference is a simple idea that sorter makers could adopt in their software. It is still a lab result on one public dataset, with no figures on line speed or yield yet.
Ziran Li of the School of Automation at Wuxi University, China, has published a new method for finding and outlining defects on potato skin in camera images. The paper appeared online in Potato Research, the journal of the European Association for Potato Research, on 7 Nov 2026 (Potato Research, Springer).
What problem does it address?
Segmentation means labeling each pixel of an image as healthy skin or defect, the step behind automated grading and post-harvest quality checks. It is hard because defects usually cover only small parts of a tuber, and healthy skin looks different from one tuber to the next. A model trained on an average idea of healthy skin can confuse normal variation with damage.
How does PotatoRelSeg work?
The model first makes a rough prediction and uses a foreground mask taken from the image to estimate what healthy skin looks like on that particular tuber. It then describes every pixel by how far it departs from that reference, using the difference, cosine similarity and Euclidean distance, alongside its original features. A small depthwise-separable residual module refines the result. No hand-drawn annotations are needed when the model is in use. During training, two extra signals push the model to align with the healthy reference and to contrast each spot with the healthy ring of skin around it.
How well does it perform?
The tests used the public Diamant Potato Dataset with its predefined split. The author reports that the method gave the best overall and defect-specific segmentation among the methods compared. Ablation tests support each component, and the model is described as compact with moderate computing overhead. The accuracy scores, model size and processing speed are in the paywalled full text and were not checked for this item.
Is it available for industry use?
The paper says processed data splits and code are available from the author on reasonable request. The work was funded by the Wuxi University Talent Start-up Fund. No trials on a commercial grading line are reported.
Sources
- PotatoRelSeg: Intra-tuber Relative Representation Learning for Lightweight Potato Surface Defect Semantic Segmentation · Potato Research (EAPR, Springer) (primary source)