2026/07/25 by Junjie Zhu, Jinpo Yang, Jian Zheng +2
Computer Science · Engineering · #Advanced Neural Network Applications #Industrial Vision Systems and Defect Detection #Modular Robots and Swarm Intelligence
paper · doi:10.1016/j.jfca.2026.109403
openalex publication_date 2026/07/25 · openalex created_date 2026/07/26 · openalex updated_date 2026/07/29
ABSTRACT Accurate hickory kernel inspection is critical for quality grading and food safety, yet manual sorting is subjective and difficult to sustain in high-throughput production. Fine-grained defect classification remains challenging because visually similar categories often differ only in subtle texture or color cues, while defect regions vary markedly in scale and morphology. To address these issues, this study proposes SHT-CNv2, an enhanced ConvNeXtV2 framework for industrial hickory kernel defect classification. The model integrates squeeze-and-excitation (SE) attention for channel-wise feature recalibration, dynamic convolution for input-adaptive representation of irregular defects, and a PANet-based multi-scale fusion head to combine shallow textural details with deep semantic information. A six-class dataset containing 14,400 images was constructed under industrial imaging conditions. SHT-CNv2 achieved a test accuracy and macro-F1 of 98.72%, with an inference speed of 133.62 FPS, outperforming representative CNN-based, Transformer-based, hybrid-attention, and lightweight architectures. Ablation studies and confusion-matrix analyses confirmed the performance gains, while Grad-CAM and t-SNE visualizations demonstrated improved defect localization and class separability. These results demonstrate the potential of SHT-CNv2 for accurate, efficient, and real-time hickory kernel sorting.