基于优化Faster R-CNN的棉花苗期杂草识别与定位

为解决棉花苗期杂草种类多、分布状态复杂，且与棉花幼苗伴生的杂草识别率低、鲁棒性差等问题，以自然条件下新疆棉田棉花幼苗期的7种常见杂草为研究对象，提出了一种基于优化Faster R-CNN和数据增强的杂草识别与定位方法。采集不同生长背景和天气条件下的杂草图像4694幅，对目标进行标注后，再对其进行数据增强；针对Faster R-CNN模型设计合适的锚尺度，对比VGG16、VGG19、ResNet50和ResNet101这4种特征提取网络的分类效果，选定VGG16作为最优特征提取网络，训练后得到可识别不同天气条件下的棉花幼苗与多种杂草的Faster R-CNN网络模型。试验表明，该模型可对杂草与棉花幼苗伴生、杂草分布稀疏或分布紧密且目标多等情况下的杂草进行有效识别与定位，优化后的模型对单幅图像平均识别时间为0.261s，平均识别精确率为94.21%。在相同训练样本、特征提取网络以及环境设置条件下，将本文方法与主流目标检测算法——YOLO算法和SSD算法进行对比，优化后的Faster R-CNN模型具有明显优势。将训练好的模型置于田间实际环境进行验证试验，识别过程对采集到的150幅有效图像进行了验证，平均识别精确率为88.67%，平均每幅图像耗时0.385s，说明本文方法具有一定的适用性和可推广性。

Identification and Localization of Weeds Based on Optimized Faster R-CNN in Cotton Seedling Stage

In order to solve the problems of low recognition rate and poor robustness of cotton seedlings and cross-growth of various weeds distribution status, seven kinds of common weeds in the field were taken as the research object under natural conditions of Xinjiang cotton seedling period. A Faster R-CNN method of growing cotton seedling weed identification with data augmentation was proposed. A total of 4694 images of weeds in cotton seedling stage under different growing backgrounds and different weather conditions were collected, then the objects of images were annotated and the data sets were augmented. The suitable anchor scale of the model was designed, and four feature extractors involving VGG16, VGG19, ResNet50 and ResNet101 were compared. VGG16 was selected as the optimal feature extractor to train cotton seedling and weeds images and optimized Faster R-CNN network detection model was obtained for weeds of different weather conditions and the variety growth status, which can effectively identify and localize seven types of weeds and cotton seedlings. The average identification time for single picture was 0.261s and the average precision of optimized Faster R-CNN was 94.21%. With the same sample, characteristic extractor network, computer condition, the proposed method was compared with the state-of-the-art methods YOLO and SSD algorithms. The results showed that the proposed Faster R-CNN model had obvious advantages in the identification of various weeds in the seedling stage of cotton field. The trained model was placed in field environment for verification test. During the recognition process, totally 150 valid images were verified, and the average recognition rate reached 88.67%. The average recognition time for each image was 0.385s. The result indicated that the proposed method had certain applicability and generalization in precise control of weeds.