Enhancing citrus surface defects detection: A priori feature guided semantic segmentation model

The accurate detection of citrus surface defects is of great importance for elevating the product quality and augmenting its market value. However, due to defect diversity and complexity, existing methods focused on parameter and data enhancement have limitations in detection and segmentation. Therefore, this study proposed a citrus surface defect segmentation model guided by prior features, named PrioriFormer. The model extracted texture features, boundary features, and superpixel features that were crucial for defect detection and segmentation, as priori features. A Priori Feature Fusion Module (PFFM) was designed to integrate the priori features, thereby establishing a priori feature branch. Then the priori feature branch was integrated into the baseline model SegFormer, with the objective of enhancing key feature learning capacity of the model. Finally, the effectiveness of the priori features in enhancing the performance of the model was demonstrated through the implementation of specific experiments. The result showed that PrioriFormer achieved an mPA (mean Pixel Accuracy), mIoU (mean Intersection over Union), and Dice Coefficient of 91.0 %, 85.8 %, and 91.0 %, respectively. Compared to other semantic segmentation models, the proposed model has achieved the best performance. The model parameters of PrioriFormer have only increase by 2.7 % in comparison to the baseline model, while the mIoU has improved by 3.3 %, indicating that the improvement of segmentation performance had less dependence on model parameters. Even when trained on few data, PrioriFormer maintained the high segmentation performance, with the reduction of mIoU not exceeding 4.2 %. This demonstrated the strong feature learning ability of the model in scenarios with limited data. Furthermore, validation on external datasets confirmed PrioriFormer's superior performance and adaptability to different tasks. The study found that the proposed PrioriFomer guided by priori features can effectively enhance the accuracy of the citrus surface defect segmentation model, providing technical reference for citrus sorting and quality assessment. • A priori feature guided semantic segmentation model PrioriFormer was proposed to enhance citrus surface defects detection. • Gabor texture features, Sobel boundary features, and superpixel region features were extracted as the priori features to improve the key feature learning ability of defect segmentation. • A Priori Feature Fusion Module (PFFM) was designed to fuse the priori features. • Comparative experiments with baseline models of different model parameters, investigation of the influence of few training data, and external validation on another dataset were conducted to verify the effectiveness of the PrioriFormer.