A novel approach for the non-destructive detection of shriveling degrees in walnuts using improved YOLOv5n based on X-ray images

Abstract

Walnuts are known for their significant nutritional and economic worth, making them the most commonly found tree nut worldwide. However, the inconsistent internal quality of walnuts reduces their market profit. Thus, their accurate classification holds significant importance in enhancing their market value. Compared with the computer vision technology, the X-ray radiography allows faster detection and more intuitive visualization of the internal conditions of the agricultural products. However, it exhibits low contrast, and thus it is difficult to define and distinguish the underlying features. Therefore, the traditional machine learning approaches are often impractical, and the deep learning-based methods are more suitable. In this study, a combination of X-ray radiography with an improved YOLOv5n model is proposed to successfully detect and classify, in real time, the shriveled walnuts into sound, slightly shriveled, moderately shriveled, and empty-shell. Due to the adopted improved YOLOv5n model, a very high classification accuracy of 97.7% can be obtained for the empty-shell walnuts, and an overall classification accuracy of 94.8% can be reached. Accordingly, the mAP, single image detection speed, and weight file size are equal to 95.78%, 6 ms, and 7.59 MB, respectively. The obtained results demonstrate that the improved light-weight YOLOv5n model allows to improve the detection accuracy for shriveled walnuts while satisfying the requirements of online real-time detection. Moreover, the X-ray radiography can be used to detect in-shell walnuts having different shriveling degrees, which also provides a reference for the internal defect detection of other dried nuts with shells.