Authors

Abstract

To address the limitations of existing detection methods for imperfect maize kernels during grain acquisition and storage — specifically limited detection categories, difficulty in identifying minute defects, insufficient robustness in individual kernel extraction, and incomplete information from single-view inspection — this paper proposes a multi-stage detection method based on CALFNet. The proposed method begins with image acquisition using a standardized imaging system, followed by a single-kernel extraction model that integrates YOLOv8 with the watershed algorithm to segment densely distributed maize kernels. The Hungarian algorithm is then employed to automatically match the front and back images of the same kernel. Based on this process, a dual-stream feature fusion classification network, CALFNet, was constructed, integrating EfficientNet-B0 and MobileNetV2 to fuse visual information from both sides of the kernels. A GUI-based detection system was subsequently developed, and the DeepSeek V3 large language model was incorporated to analyse the classification results, enabling the automatic generation of quality evaluation reports and production guidance recommendations. Experimental results show that the CALFNet model achieved a classification accuracy of 99.16% on the test set, outperforming the comparative models. In a full-pipeline integrated test on 506 real-world samples, the overall recognition and classification accuracy reached 96.05%. This study provides a feasible solution for the intelligent assessment of maize quality.

Abstract in Chinese

为解决当前粮食收储过程中玉米不完善粒检测方法存在检测类别单一、微小缺陷识别困难、籽粒单粒化提取鲁棒性不足及单面检测信息不全面等问题，本文研究并设计了一种基于CALFNet的多阶段检测方法。该方法首先通过标准化图像采集装置获取图像，并采用一种融合YOLO v8与分水岭算法的单粒化提取模型分割密集分布的玉米籽粒，再利用匈牙利算法实现同一籽粒正反面图像的自动配对。在此基础上，本文构建了名为CALFNet的双流特征融合分类网络，其基于EfficientNet-B0和MobileNet V2，旨在深度融合来自籽粒正反两面的视觉信息。基于上述方法本文搭建了GUI检测系统，并引入DeepSeek V3大语言模型对分类结果进行分析，以自动生成质量评价报告与生产指导建议。实验结果表明，CALFNet模型在测试集上的分类准确率高达99.16%，性能优于对照模型；在对506个真实样本进行的全流程集成测试中，整体识别分类准确率为96.05%。为玉米品质的智能化评估提供了解决方案。