ARTIFACT REDUCTION IN JPEG IMAGE CLASSIFICATION BY NEURAL NETWORKS: A MULTI-OBJECTIVE APPROACH
image classification, image compression, neural networks, JPEG, reduction of artifacts, multi-objective learning
Fine-grained image classification is a very important category of classification in computer vision due to its usefulness in tackling problems with a large number of classes. As well as requiring models with greater learning capacity, the datasets used for training must contain a large number of good quality samples. However lossy compression can drastically hinder the classifier’s task by degrading the signal quality. Compression and its impact on this type of application cannot be ignored, as it is something that even makes it possible to compose the training datasets for the models. In this work, we propose a Neural Network (NN) architecture capable of significantly mitigating the damage caused by JPEG compression. To do so, it relies on a double branch of structures that will be trained together. The two branches, one for Compression Artifacts Reduction (CAR) and the other for classification, are connected in such a way that the output of the first is the input of the second. The CAR branch consists of an image generation NN responsible for reducing the losses effects in the compressed signal. In turn, the classification branch uses a pre-trained NN to receive these reconstructed images as input and perform the classification. In this way, multi-objective training is able both to improve the signal quality by prioritizing important features for classification and to adapt to receiving restored images with a certain amount of degradation in order to carry out inference. In the two datasets we used, Caltech 200 Cub and Oxford 102 Flower, we increased the average accuracy for 10 different quality factors (QFs) by 45.6% and 5.81%, respectively. Despite the model’s lack of flexibility with regard to QFs and the need for additional training, our work shows a strong correlation between codec efficiency and classifier performance. It also presents an architecture that takes great advantage of this aspect to improve the compressed image classification process.