基于改进YOLOv3的田间复杂环境下菠萝拾捡识别研究

Picking recognition research of pineapple in complex field environmentbased on improved YOLOv3

为实现果实拾捡机器人在光照不均、菠萝与周围环境的颜色相似性及果实间的遮挡和重叠等田间复杂环境下对单类别菠萝的快速准确识别,提出采用深度学习下的深层残差网络改进YOLOv3卷积神经网络结构,通过单个卷积神经网络遍历整个图像,回归果实的位置,将改进的YOLOv3的3个尺度检测分别融合相同尺度模块特征层的信息,在保证识别准确率的情况下,采用多尺度融合训练网络实现田间复杂环境下端对端的单类别菠萝果实检测。最后,对改进的算法进行性能评价与对比试验,结果表明,该算法的检测识别率达到95%左右,较原始方法检测性能提升的同时,检测速度满足实时性要求,该研究为拾捡果实机器人在复杂环境下提高识别菠萝果实的工作效率和环境适应性提供理论基础。

In order to realize the rapid and accurate recognition of a single category of pineapple for the fruit picking robot in complex field environments such as uneven lighting,color similarity between pineapple and surrounding environment,and occlusion and overlap between fruits,an improve YOLOv3 convolutional neural network structure by using deep residual network was presented.And a single convolutional neural network could traverse the entire image and return to the position of the fruit,meanwhile,the improved YOLOv3 three scale detections were respectively fused with the information of the same scale module feature layer,which was used in complex environments to achieve end-to-end single-category pineapple fruit detection while ensuring recognition accuracy.Finally,the performance evaluation and comparison experiment of the improved algorithm showed that the detection and recognition rate of the algorithm was about 95%,and the detection speed met the real-time requirements while compared with the original method.It was required that the research provided a theoretical basis to improve the work efficiency and environmental adaptability of pineapple fruit recognition for the fruit picking robot in complex environments.