Training Machine Learning on JPEG Compressed Images
Maxime Pistono, Gouenou Coatrieux, Jean‐Claude Nunes, Michel Cozic · 2020
In this paper, we study the possibility to feed machine learning models with JPEG compressed images during their training phase. The underlying objective is to evaluate if JPEG decompression can be avoided so as to gain in computation time and, if yes, at what price in terms of model accuracy. To do so, we trained two well-known machine learning models: Neural Networks (NN) and Convolutional Neural Network (CNN), with pieces of data issued from different steps of the partial JPEG decompression of images. We analyze how such partially decompressed JPEG data influence machine learning model accuracy. We also study the impact of the JPEG quality factor. Experiments conducted on two image databases, MNIST database and CIFAR-10 database, show that the learning task complexity of NN models can be reduced working with partially decompressed images with a low model accuracy loss, while for CNN model's accuracy loss depends on the JPEG data. A trade-off has to be found. With a quality factor of 80, the decompression computation complexity gain is of 45% for an accuracy loss of 13%. This work also point out the need for model adapted to compressed data.