Convolutional Neural Network for Visual Artifacts Classification

Abstract

In this paper, we present an effective convolutional classifier for recognition of visual artifacts. The proposed deep-learned model is simple in architecture and number of learnable parameters while retaining a sufficient generalization ability. The latter was achieved by compilation of extensive dataset based on ImageNet database. The model was trained sequentially on over 3 million images with 3 types of distortions with various severity, specifically 10 levels of Gaussian noise, 10 levels of Gaussian blur and 5 levels of blocking effect caused by JPEG compression. The model achieved excellent 99.88% accuracy on generated images. The performance of the model was evaluated on 4 well-known IQA datasets, where it reached 71.95% accuracy in average. Furthermore, after transferring the weights from the proposed model and short training on IQA datasets, its accuracy increased by more than 7%.