激光与光电子学进展, 2020, 57 (12): 121007, 网络出版: 2020-06-03
改进Faster RCNN模型在棉花异性纤维识别中的应用 下载: 807次
Application of Improved Faster RCNN Model for Foreign Fiber Identification in Cotton
图像处理 异性纤维 深度学习 目标识别 Faster RCNN k-means++ image processing foreign fiber deep learning object detection Faster RCNN k-means++
摘要
采用深度学习方法对棉花中的异性纤维进行分类识别。首先建立异性纤维数据集,针对异性纤维尺寸和形状多样性的特点,采用基于Faster RCNN的目标识别框架,以RseNet-50代替原始的VGG16作为异性纤维分类模型的特征提取网络,并采用k-means++聚类算法对候选框生成尺寸进行改进;然后对模型进行训练,实现棉花中异性纤维的分类和定位。训练后的模型在验证集上的准确率达到94.24%,精度为98.16%,召回率为95.93%,精确率和召回率的调和平均数(F1分数)为0.970。对比改进前、后模型对异性纤维的识别效果,改进后的模型在小尺寸、大长宽比和密集出现的情况具有更好的识别效果,相对于原始模型,其准确率、精度、召回率和F1分数分别提高了3.21%、0.90%、2.51%和0.017。
Abstract
Generally, deep learning methods are used to identify and classify the foreign fibers in cotton. First, a target recognition framework is adopted based on Faster RCNN to develop a foreign fiber dataset according to the characteristics of foreign fiber size and shape diversity. Next, the original VGG16 is replaced by RseNet-50 as the feature extraction network in the foreign fiber classification model, and the size of the mark box is improved using the k-means++ clustering algorithm. Subsequently, the model is trained to identify and classify the foreign fibers in cotton. The trained model achieves an accuracy rate of 94.24%, a precision of 98.16%, a recall rate of 95.93%, and an F1 score of 0.970 with respect to the verification set. When compared with the original model, the recognition effect is observed to improve in case of small sizes, large aspect ratios, and dense occurrences when the proposed model is used. Furthermore, the accuracy, precision, recall rate, and F1 score of the proposed model improve by 3.21%, 0.90%, 2.51%, and 0.017, respectively, when compared with those of the original model.
杜玉红, 董超群, 赵地, 任维佳, 蔡文超. 改进Faster RCNN模型在棉花异性纤维识别中的应用[J]. 激光与光电子学进展, 2020, 57(12): 121007. Yuhong Du, Chaoqun Dong, Di Zhao, Weijia Ren, Wenchao Cai. Application of Improved Faster RCNN Model for Foreign Fiber Identification in Cotton[J]. Laser & Optoelectronics Progress, 2020, 57(12): 121007.