Xception-based architecture with cross-sampled training for Image Quality Assessment on KonIQ-10k

Abstract

Image quality assessment is a crucial task in various fields such as digital photography, online content creation, and automated quality control, as it ensures an optimal visual experience and aids in maintaining consistent standards. In this paper, we propose an efficient method for training image quality assessment models on the KonIQ-10k dataset. Our novel approach utilizes a dual-Xception architecture that analyzes both the image content and additional image parameters, outperforming traditional single convolutional models. We introduce cross-sampling methods with random draw sampling of instances from majority classes, effectively enhancing prediction quality in the Mean Opinion Score (MOS) ranges that are underrepresented in the database. This methodology allows us to achieve near state-of-the-art results with limited computing costs and resources. Most importantly, our predictions across the entire spectrum of MOS values maintain consistent quality. Because of using a novel and highly effective method for image sampling, we achieved these results with much lower computational cost, making our approach the most effective way of MOS estimation on the KonIQ-10k database.