一种基于可见光巡检图像的杂草智能识别方法

An Intelligent Weed Recognition Method Based on Optical Patrol Image

目前电力巡检主要是采用无人机巡检的方式,针对无人机巡检获取的图像识别过程中,电力设备旁的杂草可能会造成安全隐患,需要对图像中的杂草进行识别。针对电力巡检的场景,提出了一种基于可见光巡检图像的杂草智能识别方法,以可见光巡检图像中杂草的特征为基础,结合卷积神经网络方法,解决可见光巡检图像中电力设备附近的杂草识别问题。通过对图像进行样本数据增广和预处理,接着引入区域生成网络,再对图像提取固定个数候选框的图像特征,和改进的图像分类网络连接在一起,得到最终的卷积神经网络模型。实验表明其准确率可以达到97.98%,检测一幅600×600大小图像需要花费的平均时间约为0.256 s,在保证了准确率的同时达到了高效识别的要求。

Power inspection is currently carried out mainly by drones. The weeds around the power equipment may cause potential safety hazards when the patrol images acquired by drones are used for patrol inspection, it is therefore necessary to recognize the weeds in the image. In this paper, a method for intelligent recognition of weeds is proposed for power patrol inspection based on optical patrol images. Based on the feature of weeds in the optical images, and combined with the convolutional neural network method, the problem of weed recognition near the power equipment in optical patrol images is solved. By amplifying and preprocessing the sample data of the optical patrol images, and introducing the region proposal network, the image features of the fixed number of candidate frames are extracted from the images. Then the network is connected to the improved image classification network to obtain a final convolutional neural network model. The experiments show that the accuracy rate can reach 97.98%, and the average time taken for detecting a 600×600 image is around 0.256 seconds, which meets the requirements of efficient recognition while ensuring accuracy.