Murgio/Food-Recipe-CNN

Combines transfer learning from pre-trained CNNs (AlexNet, VGG, ResNet, GoogLeNet) with approximate nearest-neighbor search using PCA-reduced feature vectors to match food images against 400K+ images and retrieve recipes from a 300K+ German-language recipe dataset. The hybrid approach compares CNN category predictions with the top-5 optically similar images to improve recipe retrieval accuracy, addressing the challenge of visually similar dishes across different food categories. Built with Keras and deployed as DeepChef, a web application accepting meal photos as input.

585 stars. No commits in the last 6 months.