基于卷积神经网络的食品图像识别

Application of convolutional neural network in image retrieval and classification of food image

中文关键词:  食品图像  图像检索  图像分类  Inception_V3-CNN  卷积神经网络

英文关键词:food image  image retrieval  image classification  inception_V3-CNN  convolutional neural network

中文摘要:

      目的 探究一种基于Inception_V3-CNN模型的食品图像识别和分类方法。方法 选取包含20类食品和19609张的食品图像建立数据集Food-101, 通过卷积神经网络(convolutional neural networks, CNN)提取图像候选区域的视觉特征, 并自动进行分类, 使其具有较高的识别率; 此外, 采集来自kaggle库中已标注的食品图像集做预测对比实验, 以保证Inception_V3-CNN模型检测的准确度。结果 该方法能够去除背景噪音, 且仅需部分提取视觉特征, 可以有效解决网络食品图像的分类问题, 与多视图支持向量机(support vector machine, SVM)+方向梯度直方图(histogram of oriented gradients, HOG)和传统CNN方法相比, 在测试时间相近、设备计算能力相同的条件下, 该方法识别率更高, 迭代次数为15000次时, Loss值降至4.92, 准确率可达93.89%。结论 此方法可以快速识别食品图像, 在实际网络图片中能有较好的可移植性。将算法移植到移动设备中实现APP的模块化操作也将成为后续工作探索的重点方向。

英文摘要:

      Objective To investigate a food image recognition and classification method based on Inception_V3-CNN model. Methods The food image dataset Food-101 containing 20 foods and 19609 food images was established, and the visual features of the candidate regions of the image were extracted by convolutional neural network (CNN), and automatically classified to have a higher recognition rate. The food image set from the kaggle library was collected for predictive comparison experiments to ensure the accuracy of the Inception_V3-CNN model. Results This method could remove background noise and only needed to extract visual features in part, which could effectively solve the classification problem of network food images. Compared with support vector machine+ histogram of oriented gradients (SVM+HOG) and traditional CNN methods, under the conditions of similar test time and the same computing power of equipment, the recognition rate of this method was higher. When the number of iterations was 15000 times, the Loss value dropped to 4.92, and the accuracy rate could reach 93.89%. Conclusion This method can quickly recognize food images and has good portability in real network images. Transplantation of the algorithm to mobile devices to realize the modular operation of APP will also become the focus of future work.