基于改进YOLO 11的轻量化鹌鹑蛋裂纹在线无损检测系统

Lightweight Quail Egg Crack Online Non‑destructive Testing System Based on Improved YOLO 11

摘要:

针对鹌鹑蛋裂纹缺陷视觉无损检测效率低、准确性不足的问题，本文提出一种基于改进YOLO 11的轻量化鹌鹑蛋裂纹在线无损检测系统。以YOLO 11为基线模型，通过引入 C3k2_IDWC、SPPF_LSKA 及 EUCB_SCM 模块完成网络结构优化，再经 LAMP 剪枝实现全局稀疏化，并结合BCKD逻辑蒸馏与CWD特征蒸馏迁移判别知识，构建兼具高精度与轻量化特性的 YOLO 11?Quail 模型。试验结果表明，改进的YOLO 11?Quail模型平均精度均值达到94.6%，精确率和召回率均为95.1%，检测速度达184.8 f/s，参数量仅为1.6×10^6。在Jetson Nano部署测试中，采用数字化结果矩阵判定跟踪算法实现多角度综合检测，误检率降低至8.33%，检测时延为68.59 ms，较改进前处理速度提升11.3%，满足典型鹌鹑蛋裂纹在线无损检测需求。

Abstract:

Aiming to address the critical quality control challenge of detecting cracked quail eggs in egg processing, where manual and machine?assisted visual inspection are both inefficient and inaccurate, a high?performance, lightweight online non?destructive detection system was developed based on a comprehensively improved YOLO 11 deep learning architecture. YOLO 11n was selected as the baseline for its favorable balance of speed and accuracy, then systematically optimized for the specific task of identifying subtle cracks on variably patterned quail eggshells through two core phases: network architectural refinement and model compression. In the network refinement phase, the backbone network integrated the C3k2_IDWC module to capture multi?scale textural details and directional contextual information of fine cracks, while the neck network replaced the original SPPF module with SPPF_LSKA (to focus on crack regions via spatial attention) and standard upsampling with EUCB_SCM (to enhance cross?channel information interaction and feature fusion). For model compression, the enhanced network underwent global sparsification via the LAMP pruning algorithm to eliminate redundant parameters, followed by dual knowledge distillation combining BCKD (logit?level guidance) and CWD (feature?level imitation) to transfer knowledge from the high?performance teacher model to the compact student model, yielding YOLO 11?Quail. Experimental results on a dark?field backlighting dataset showed YOLO 11?Quail achieved 94.6% mAP, 95.1% precision and recall, 184.8 f/s inference speed, and only 1.6×10^6 parameters. Deployed on Jetson Nano with a multi?angle tracking algorithm, it delivered an 8.33% false detection rate, 0 missed detection rate, and 68.59 ms latency, satisfying the high?precision real?time demands of industrial production lines.