本文为芬兰阿尔托大学（作者：Janaki Prasad Koirala）的硕士论文，共64页。

识别食物图像可以拯救人们的生命。它可以用来了解食物中潜在过敏原的存在，或者估计食物的营养含量，也可以用来对抗肥胖的流行。考虑到这些应用，我们试图利用机器学习和深度学习的进展来训练从数码照片中识别欧洲食品的模型。从研究文献中发现，速度较快的RCNN是目前最先进的基于CNN的图像局部信息获取和识别框架。此外，我们还开发了一个Android应用程序来识别食物对象。更快的RCNN需要大量的数据，其中包含对象的标签和定位信息。要找到这样的数据集来训练我们的网络是非常困难的。我们用445个标签制作了一个69k图像的食物数据集，并用这些图像训练我们的模型。但是数据集在每个类别的图像数量方面存在偏差，这会对模型的性能产生负面影响。为了提高性能，我们尝试了几种方法，比如只提取标签的子集，并为每个标签均衡训练样本的数量。在训练样本有限的情况下，我们也使用迁移学习来解决网络过度拟合的问题。最后，通过使用公开的数据集并根据我们的需要进行调整，我们的模型能够以0.37的平均精度识别图像。Android应用程序使用这个模型从图像中识别食物对象。

Identifying a food from its image can savepeople’s life. It can be used to know the presence of potential allergens infood or by estimating the nutritional content of food, it may also be used tocombat the obesity epidemic. With such applications in mind, we seek to exploitthe advances in machine learning and deep learning to train models thatidentify European food from digital photos. From the literature it wasdiscovered that the Faster RCNN was the current state-of-art CNN basedframework which could get local information of object in image and recognizeit. Furthermore, we also develop an Android application for recognition of foodobjects. Faster RCNN requires a large volume of data with labels andlocalization information of the objects present in them. It is very challengingto find such datasets to train our network. We made up a food dataset of 69kimages with 445 labels and trained our model using those images. But thedataset was skewed in terms of numbers of images per category that negativelyaffected the performance of the model. To improve the performance, we triedseveral approaches like taking only a subset of labels and equalizing thenumber of training samples for each label. We also used transfer learning to getaround the problem of overfitting the network when our training sample size islimited. Finally, by using publicly available data set and adapting it to ourneeds, our model was able to identify images with 0.37 mean Average Precision.The Android application uses this model to recognize food objects from images.

1 引言

2 项目背景

3 目标识别

4 实验

5 结果

6 结论

附录 排行榜前90的食品类别

更多精彩文章请关注公众号：