Abstract: In order to solve the problem of low efficiency and high subjectivity of traditional wheat identification, an improved CBAMInceptionV3 wheat impurity identification method was proposed. Firstly, a machine vision online detection platform was built to collect dynamic image data, and the wheat impurity image was processed by data set enhancement, image preprocessing and KS classification. Then, GoogLeNet, ResNet34 and InceptionV3 models were used to classify and train the image data set. Secondly, based on InceptionV3 model, CBAM was introduced to enhance the sensitivity of the model to information and improve the recognition accuracy of the model. The improved convolutional neural network CBAMInceptionV3 model is compared with CAInceptionV3 and InceptionV3 models added in CA module. The results show that the accuracy of InceptionV3 model on test set is 83.5% and F1Score is 82.41%, and the accuracy of CAInceptionV3 model on test set is 92.3% and F1Score is 92.29%. CBAMInceptionV3 has 92.9% accuracy and 92.92% F1Score on the test set. The average prediction time of CBAMInceptionV3 model for the test set is 0.045 pieces/s, which is significantly better than the other two models.

Key words: wheat impurity, convolutional neural network, classification and recognition, CBAMInceptionV3, visualization

摘要： 为解决传统人工识别小麦效率低、主观性高的问题，提出一种基于改进的CBAMInceptionV3小麦杂质识别方法。搭建机器视觉在线检测平台采集动态图像数据，采用数据集增强、图像预处理和KS分类方法对小麦杂质图像进行处理；选用GoogLeNet、ResNet34、InceptionV3三种模型对图像数据集进行分类训练。以InceptionV3模型为基础，引入注意力机制CBAM，增强模型对信息的敏感度，提升模型的识别准确率。将改进卷积神经网络CBAMInceptionV3模型与加入CA模块的CAInceptionV3、InceptionV3两种模型进行对比试验。结果表明，InceptionV3模型在测试集上准确率为83.5%、F1Score为82.41%，CAInceptionV3模型在测试集上准确率为92.3%、F1Score值为92.29%，CBAMInceptionV3在测试集上准确率为92.9%、F1Score值为92.92%。CBAMInceptionV3模型对测试集的平均预测时间为0.045张/s，明显优于其他两种模型。

关键词: 小麦杂质, 卷积神经网络, 分类识别, CBAMInceptionV3, 可视化