A method of citrus epidermis defects detection based on an improved YOLOv5. (March 2023)

• Record Type: Journal Article
• Authors: Hu, WenXin; Xiong, JunTao; Liang, JunHao; Xie, ZhiMing; Liu, ZhiYu; Huang, QiYin; Yang, ZhenGang
• Is Part Of: Biosystems engineering. Volume 227(2023)
• Page Start: 19
• Page End: 35
• Publication Date: 2023-03
• Subjects: Convolutional neural network -- Attention mechanism -- Loss function -- Intelligent detection -- Fluorescent; MLP Multilayer Perceptron; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2023.01.018
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : Achieving intelligent detection of citrus epidermal defects after harvesting is of great significance to citrus quality and value assurance. The uneven illumination makes it challenging to detect citrus epidermis abnormalities with great accuracy. In order to improve the detection accuracy of citrus epidermal invisible defects, firstly, a dual-lamp image acquisition system is designed and used to complete the image acquisition of citrus fruit invisible defects. Secondly, the YOLOv5 model was optimized by integrating the attention mechanism CBAM and modifying the loss function as DIoU. Finally, the performance of the improved model was verified by comparison experiments and ablation experiments. According to the experimental results, mAP, Precision and Recall of the improved YOLOv5 model were 95.5%, 94.0% and 95.1%, respectively, which were 5.8%, 3.6% and 7.6% higher than those of YOLOv5x. Meanwhile, the average detection speed increased by 22.1 ms per pic. This indicates that the improved network being applied to citrus epidermal defects detection can achieve better performance. It can provide technical support for the intelligent detection and grading of postharvest citrus. Highlights: A YOLOv5 method of citrus epidermis defects detection was proposed. Automatic and rapid detection of invisible defects in citrus epidermis was realised. The optimisation of attention mechanism and loss function improved the network. Fluorescent defect images were obtained by dual illumination image acquisition.