FDE-RCF: A Navel Orange Defect Edge Detection Algorithm Based on Feature Depth Enhancement

A new edge detection method is proposed for the inspection of navel orange surface defects, enabling precise defect localization, classification type, and size estimation. This study facilitates accurate and efficient grading and sorting, thereby ensuring product quality and extending storage life. To overcome the limitations of existing edge detection methods in handling small, blurred, and color-similar defects, a novel Feature Depth-Enhanced Edge Detection Model (FDE-RCF) is introduced. This model enhances the Richer Convolutional Features (RCF) network by removing its final pooling layers to preserve fine-grained details and integrating three key modules: (1) Dilated Convolution for improved global context, (2) Efficient Channel Attention (ECA) module to selectively focus on defect-relevant features, and (3) Scale Enhancement Module (SEM) for robust edge detection across varying object sizes. A residual fusion structure is also introduced to better capture intricate edge patterns. Experimental results demonstrate that FDE-RCF outperforms conventional models. On navel orange defect datasets, it achieves higher ODS (83.04%), OIS (84.97%), and AP (86.64%), with improvements of 2.32%, 2.44%, and 2.48% respectively, over the baseline RCF. On the standard BSDS500 benchmark, it achieves additional ODS and OIS by 1.81% and 1.78%, while maintaining real-time processing speeds at 28 FPS. The proposed FDE-RCF model generates coherent and continuous edge representations and accurately capture subtle and small-scale defects, highlighting its suitability for integration in automated fruit grading and visual quality inspection systems.