基于深度学习的杂草识别系统

田间除草技术在农业生产中具有重要意义。针对复杂背景下农作物与杂草识别率低、算法鲁棒性差等问题，提出一种图像分割网络 Res-Unet。该网络为 unet 网络的改进版本，采用 resnet50 网络代替 unet 主干网络，解决复杂背景下农作物与杂草区域提取困难、小植株检测效果差、分割边缘震荡、变形问题。将图像的平均交并比、准确率、训练时长作为评价指标进行实验。结果表明：使用 Res-Unet 模型的平均交并比为 82.25%，平均像素准确率为 98.67%。改进的 Res-Unet 模型相对于 Unet 平均交并比高出 4.74%，相较于 segnet 平均交并比高出 10.68%，训练时间减少 3 小时。该方法对复杂背景下甜菜杂草检测效果良好，可为机器人精确除草提供参考。

Weed Identification System Based on Deep Learning

The field weeding technology is of great significance in agricultural production. The traditional weed identification technology has the disadvantages of low efficiency or great limitations. To solve the problem of low recognition rate of crops and weeds in complex background and poor robustness of algorithm，an image segmentation network res UNET is proposed. Res UNET is an improved version of UNET network. It uses resnet50 network instead of the main network of UNET to solve the problem of crop and weed area extraction under complex background，poor detection effect of small plants，edge vibration and deformation of segmentation. The average intersection ratio，accuracy and training time of the image are selected as evaluation indexes. The results show that the average cross union ratio of res UNET model is 83.25%，and the average pixel accuracy is 98.67%. The improved res UNET model is 4.74% higher than the UNET average，10.68% higher than the segnet average，and the training time is reduced by 3 hours. This method has a good detection effect on beet weeds in complex background，and can provide a reference for the follow-up robot precision weeding.