An Impurity Rate Estimation Method of Post-Harvest Sugarcane Based on Rotated Bounding Box and Binocular Vision

Abstract

Sugarcane is an important economic crop. After machine harvesting, the impurity rate of sugarcane is an important metric, which affects the sugar output rate. In order to obtain the impurity rate and detect primary impurities, this paper proposes an impurity rate estimation method for post-harvest sugarcane based on rotated bounding box and binocular vision. Firstly, the sugarcane mixture image including sugarcane segments, sugarcane tops, and sugarcane leaves was captured by a binocular camera. Secondly, we replaced the Darknet53 backbone with the EfficientNetV2 network in the improved YOLOv5-obb algorithm to obtain the rotated bounding boxes for the sugarcane mixture. Next, based on binocular vision, the dimensions of sugarcane segments are calculated. And the sugarcane segments are fitted as cylinders, enabling the calculation of their volume and mass. Finally, the impurity rate of post-harvest sugarcane is calculated based on the mass of sugarcane segments and the total mass of the mixture. Experimental results show that rotated bounding boxes can fit the shape of each target accurately, with a mean average precision (mAP) of 97.6%. The model also effectively detects occluded and overlapping targets, while reducing parameters by 36% compared to YOLOv5-obb. The average detection time per image is 0.017 s, and the average time for impurity rate estimation per image is 0.19 s. For 830 test images, the average mass error of sugarcane segments is 10.31%, the total mass error is 0.23%, and the total impurity rate error is 4.58%.

Keywords

• Sugarcane
• rotated bounding box
• binocular vision
• impurity rate
• machine vision