基于卷积神经网络的多尺度葡萄图像识别方法

葡萄品种质量检测需要识别多类别的葡萄,而葡萄图片中存在多种景深变化、多串等多种场景,单一预处理方法存在局限导致葡萄识别的效果不佳。实验的研究对象是大棚中采集的15个类别的自然场景葡萄图像,并建立相应图像数据集Vitis-15。针对葡萄图像中同一类别的差异较大而不同类别的差异较小的问题,提出一种基于卷积神经网络(CNN)的多尺度葡萄图像识别方法。首先,对Vitis-15数据集中的数据通过三种方法进行预处理:旋转图像的数据扩增方法、中心裁剪的多尺度图像方法以及前两种方法的数据融合方法;然后,采用迁移学习方法和卷积神经网络方法来进行分类识别,迁移学习选取ImageNet上预训练的InceptionV3网络模型,卷积神经网络采用AlexNet、ResNet、InceptionV3这三类模型;最后,提出适合Vitis-15的多尺度图像数据融合的分类模型MS-EAlexNet。实验结果表明,在同样的学习率和同样的测试集上,数据融合方法在MS-EAlexNet上的测试准确率达到了99.92%,相较扩增和多尺度图像方法提升了近1个百分点,并且所提方法在分类小样本数据集上具有较高的效率。

Multi-scale grape image recognition method based on convolutional neural network

Grape quality inspection needs the identification of multiple categories of grapes, and there are many scenes such as depth of field changes and multiple strings in the grape images. Grape recognition is ineffective due to the limitations of single pretreatment method. The research objects were 15 kinds of natural scene grape images collected in the greenhouse, and the corresponding image dataset Vitis-15 was established. Aiming at the large intra-class differences and small inter-class of differences grape images, a multi-scale grape image recognition method based on Convolutional Neural Network (CNN) was proposed. Firstly, the data in Vitis-15 dataset were pre-processed by three methods, including the image rotating based data augmentation method, central cropping based multi-scale image method and data fusion method of the above two. Then, transfer learning method and convolution neural network method were adopted to realiize the classification and recognition. The Inception V3 network model pre-trained on ImageNet was selected for transfer learning, and three types of models-AlexNet, ResNet and Inception V3 were selected for convolution neural network. The multi-scale image data fusion classification model MS-EAlexNet was proposed, which was suitable for Vitis-15. Experimental results show that with the same learning rate on the same test dataset, compared with the augmentation and multi-scale image method, the data fusion method improves nearly 1% testing accuracy on MS-EAlexNet model with 99.92% accuracy, meanwhile the proposed method has higher efficiency in classifying small sample datasets.