Food classification using UPMC Food-101 dataset
A dataset containing almost 100,000 images of 101 dishes collected from recipes websites on Internet. The dataset also contains the textual data from these website. For more information, please have a look at :
http://visiir.lip6.fr/explore
We try to classify the different dishes by fine-tuning the pre-trained models : VGG16, Resnet18, InceptionV3.
We also try to classify the dishes by the textual features collected from the websites, we then do a fusion of the image classification and textual classification.
After using several methods of data processing (data augmentation, rescale), we obtain the following results :
- VGG16 : 0.5242
- Resnet18 : 0.6667
- InceptionV3 : 0.7217
For the textual classification, we obtain an accuracy of 0.83 for a neuron network, 0.87 for the SVC Linear.
The result after combining image and texutal features : 0.8836.
More details on the data as well as results is presented in the file: Projet MAP 583.pdf.