基于改进YOLOX模型的樱桃缺陷及分级检测

目的:实现工业化条件下樱桃的快速分级。方法:采用YOLOX网络对缺陷果进行检测,通过为特征金字塔网络设置适当的融合因子来提高不明显缺陷的检测精度,并将Focal Loss集成到损失函数中;使用YOLOX网络对完好果进行分级,引入注意力机制CBAM来加强网络特征提取。结果:樱桃表面缺陷的平均检测精度为97.59%,大小和颜色分级的平均检测精度为95.92%。结论:改进后的YOLOX网络可明显提升樱桃缺陷及分级检测的精度。

Cherry defect and classification detection based on improved YOLOX model

Objective: In order to expand the scope of cherry sales and achieve rapid grading of cherries under industrial conditions. Methods: Firstly, the YOLOX network was used to detect the defective fruit, in order to solve some problems where the defect was not obvious. The detection accuracy of the inconspicuous defect was improved by setting the appropriate fusion factor for the feature pyramid network, and in order to solve the problem of imbalance between various types of real samples, Focal Loss was integrated into the loss function. Then, the intact fruit was graded using the YOLOX network, and the attention mechanism CBAM was introduced to enhance the network feature extraction. Results: Experimental results showed that 97.59% of the mAP detected for cherry surface defects and 95.92% of the mAP of size and color grading. Conclusion: The accuracy of cherry defects and grading has been significantly improved by the improved YOLOX network.