Authors

Abstract

This study addressed the practical problems of complex picking environments, difficult image recognition, and low picking efficiency in apple harvesting, combined with China's agricultural requirements and picking systems. An improved apple fruit recognition method based on attention mechanisms and YOLOv5 was proposed. A dataset was created by collecting 3,600 apple images under front-light, side-light, and backlight conditions at different coloring stages in natural environments. The SENet and CBAM attention mechanisms were used to enhance YOLOv5's feature extraction network, and the model was trained to improve detection accuracy. Experimental verification showed that the YOLOv5x model embedded with the CBAM module achieved the highest mean average precision (mAP) of 98.3%. The CBAM module outperformed the SENet module. Actual tests of the apple-picking robot's vision system prototype showed that when the IOU threshold was set at 0.5 and 0.3, the average detection accuracy was over 85% in both cases. The results demonstrated that the improved YOLOv5 model exhibited robustness to light intensity variations. This approach provides a technical reference for developing apple picking robot vision systems.

Abstract in Chinese

针对苹果采摘存在采摘环境复杂、图像准确识别困难、采摘效率低下等实际问题，结合我国苹果采摘农艺要求及采摘体系。本文提出了一种基于注意机制和改进的YOLOv5的苹果果实识别方法。该方法通过收集3600张自然环境中顺光、侧光和背光的不同着色天数的苹果图像，创建了一个数据集，注意机制SENet和CBAM用于改进YOLOv5的特征提取网络，并对模型进行训练以提高模型的检测精度。经过实验验证，嵌入CBAM模块的YOLOv5x的平均检测精度最高，mAP为：98.3%。CBAM模块的性能优于SENet模块。结果表明：改进的YOLOv5模型对光强变化具有良好的鲁棒性。通过采摘机器人视觉识别系统样机的实际试验验证，当IOU阈值设为0.5和0.3时，该系统样机平均检测精度均在85%以上。改进后的YOLOv5模型可为苹果采摘机器人视觉系统的开发提供参考。