Abstract: In order to identify the location of Hyphantria cunea larvae nets more quickly and accurately, the spraying robot was controlled to spray the target, in this paper, a fast recognition algorithm based on chromatic difference threshold segmentation method and improved convolutional neural network (CNN) was proposed for the screen image of American white moth larvae. Firstly, the threshold segmentation based on color difference was carried out on the image, the probability of containing the screen was preliminarily determined, and the qualified image with the probability higher than a certain value was determined as the qualified image. The qualified image was sent to the convolution neural network, the image was non coincident traversed, the screen outline frame was marked, and further judge whether the optimal spraying position was reached. When the optimal spraying position was reached, the system stopped moving for spraying, when the image was unqualified or did not reach the best spraying position, the program controlled the actuator to continue to move according to the planned walking path. The experiment showed that the accuracy of screen recognition was more than 95%, the image recognition time was less than 200 ms, and there was no leakage in the experiment.

Key words: Hyphantria cunea larvae nets, threshold segmentation, convolutional neural network, visual servo

摘要： 为更快速、更精确识别美国白蛾幼虫网幕位置，控制喷药机器人进行对靶喷药，以美国白蛾幼虫网幕为研究对象，提出一种基于色差的阈值分割方法与改进卷积神经网络(CNN)相结合的网幕图像快速识别算法并对其视觉伺服系统进行研究。对图像进行基于色差的阈值分割，初判其含有网幕的概率，将概率高于一定值的定为合格图像，将合格图像送入卷积神经网络，对图像进行不重合遍历，标出网幕轮廓框，并进一步判断是否达到最佳喷药位置，达到最佳喷药位置时系统停止运动进行喷药，当图像不合格或未达到最佳喷药位置时由程序控制执行机构按照规划好的行走路径继续运动。试验表明：网幕识别的准确率达到95%以上，图像的识别时间在200 ms以内，试验无漏喷。

关键词: 美国白蛾幼虫网幕, 阈值分割, 卷积神经网络, 视觉伺服