基于迁移学习和批归一化的菜肴图像识别方法

FOOD IMAGE RECOGNITION BASED ON TRANSFER LEARNINGAND BATCH NORMALIZATION

菜肴图像识别属于图像细粒度识别。针对菜肴子类之间差距小、外观差异大且受外界因素影响难以识别问题,提出一种基于迁移学习和批归一化结合的深度学习模型菜肴图像识别方法。以预训练的VGG-16为迁移学习基础,对部分卷积层以及全连接层输出做批归一化处理,最终得到尺度变换和平移后的特征集合。通过迁移学习解决深度学习所带来的过拟合问题,获取比人工特征更具有鉴别性的深度特征;通过批归一化处理缓解深度学习中存在的梯度消失问题。迁移学习的相关实验中以loss、top1、top5准确率为指标;批归一化相关实验中以top1准确率和top5准确率为指标。实验表明,在VireoFood172和UEC-Food100数据集上,所提出的模型与原始模型相比,loss明显下降,准确率有大幅提升,并且与现有方法相比在菜肴图像识别的top1和top5准确率上均有所提升。

Food image recognition is a kind of fine-grained image recognition.Considering small gaps among subclasses of various food,large differences in appearance and other uncertain external factors make it difficult to recognize food images,a deep learning model based on transfer learning and batch normalization is put forward to deal with these problems.Based on the pre-trained VGG-16 model,outputs of partial convolution layers and all fully connected layers were normalized,and we obtained the features after scale transform and scale translation.Transfer learning was applied to the model to overcome over-fitting caused by deep learning in some way as well as obtaining more discriminative in-depth features than artificial features.Batch normalization could help solve the problem of gradient disappearance in deep learning.The indicators in the related experiments of transfer learning were loss,top1 precision and top5 precision,while top1 precision and top5 precision were used as indicators in experiments related with batch normalization.The results of the experiments show that the loss decreases significantly,and the precision is greatly improved on VireoFood 172 and UEC-Food 100 datasets compared with the primitive model.Compared with the existing methods,the accuracy of top 1 and top 5 of food image recognition is improved.