Yakiniku Doneness Degree Recognition System using Autoencoder and Convolutional Neural Network

This paper presents a doneness recognition system using visual features on meat slices in the Japanese barbeque or ‘yakiniku’ cooking style. The recognition system employs a simple convolutional neural network (CNN) and an autoencoder as the main components. The autoencoder is applied to reduce two major distortions in input images, including image temperatures and blurs. The CNN recognizes optimized images from the autoencoder as one of three classes, which are overcooked, undercooked, and well-cooked. The proposed system is evaluated using distorted images through two experiments: the experiment using the CNN individually, and the experiment using the autoencoder and the CNN. Results from experiments are evaluated through precisions, recalls, and F1-scores. Experimental results depict satisfying recognition performances of the system with a CNN and an autoencoder on images that simulate inputs from bad visual devices.