LIGHTWEIGHT YOLOV5S-SUPER ALGORITHM FOR MULTI-DEFECT DETECTION IN APPLES

Abstract

ABSTRACT As the application scenarios of embedded devices become increasingly extensive, the use of high-performance convolutional neural networks can solve the problem of low accuracy of multiple defects detection in apples. However, owing to the overly large parameters and network structure of the convolutional neural network, perfectly integrating it with the embedded devices is difficult. Therefore, this study proposes a lightweight and improved algorithm based on Yolov5s. First, the structure of the optimized MobileNetV3 is introduced in the backbone layer to reduce the computational and parametric quantities of the model. Wise-IoU is used as the loss function of the localization regression of the bounding box to reduce the harm of low-quality samples on anchor box regression. The efficient multiscale attention mechanism is embedded in each downsampling layer of the backbone, and small target detection is added to the neck layer to improve the attention of the convolutional layer on important features. The experimental results showed that the Yolov5s-Super model parametric count decreased by 78%, and accuracy P, mAP@50, and mAP@50:95 improved by 10.3%, 3.2%, and 4.2%, respectively, compared to the original model. Theoretical support is provided for the migration of this network model to embedded devices.

Keywords

• apple defect
• Yolov5s-Super
• lightweighting
• EAM attention
• Wise-IoU