基于迁移学习的竹片缺陷识别

使用竹片图像实现竹片缺陷自动识别,目前深度学习可以有效地解决该类问题,但是必须使用大量样本数据做训练才能获得较高的识别准确率。当图像数量有限时,利用基于迁移学习的方法,把经过预训练的卷积神经网络模型进行迁移,即共享卷积层和池化层的权重参数,调整新网络模型的超参数,并建立一个包含4种共计6 360张竹片缺陷图像的数据库,把图片分成4种训练集测试集形式,即80%训练、20%测试;60%训练、40%测试;40%训练、60%测试;20%训练、80%测试,分别利用支持向量机SVM分类方法、深度学习方法和迁移学习方法进行训练和测试,并将这3种方法作对比。最后,通过构建竹片缺陷识别的混淆矩阵对迁移学习进行具体分析与说明。结果表明,按照80%训练、20%测试的识别准确率最高,通过迁移学习得到的竹片缺陷最高识别精度分别达到98.97%,比普通深度学习提高了11.55% ,比SVM分类方法提高了13.04%。说明迁移学习比普通深度学习和传统支持向量机SVM分类方法更适合用于小样本数据集的分类识别,并且效果优于普通深度学习和 SVM 分类方法。

Defect Recognition of Bamboo Sheets Based on Transfer Learning

At present,deep learning can effectively realize the automatic identification of the defects of bamboo sheets,in which however,a large number of sample data must be used for training in order to achieve high recognition accuracy.To solve the problem of limited images,we tried to use a new method that was based on the migration of learning to achieve the defect identification of bamboo sheets.In this method,a the migration operation was conducted for the preliminarily trained convolution neural network model,namely,to share the weight of convolution and pooling layer parameters,and to adjust the new super network model parameters.A database that included four types of defected bamboo sheet images with the number of 6 360 were established.The pictures were divided into four training and testing sets,namely,80% (training) & 20% (testing),60%&40%,40%&60%,20%&80%,the four types were trained and tested by three methods:support vector machine SVM classification method,the deep learning method,and migration study methods.The results by three methods were compared.Finally,the transfer learning was analyzed and explained concretely by constructing the confusion matrix of bamboo sheet defect recognition.The experimental results showed that 80% training &20% testing had the highest recognition accuracy.The highest recognition accuracy of bamboo sheet defects obtained by transfer learning was 98.97%,11.55% higher than the deep learning method and 13.04% higher than the SVM classification method.It showed that the transfer learning was more suitable for the classification and recognition of small sample data sets than the common deep learning and the traditional support vector machine SVM classification methods,and the results were better than the common deep learning and SVM classification methods.