Title: Perceptual Quality Assessment for Fine-Grained Compressed Images

Abstract: Recent years have witnessed the rapid development of image storage and transmission systems, in which image compression plays an important role. Generally speaking, image compression algorithms are developed to ensure good visual quality at limited bit rates. However, due to the different compression optimization methods, the compressed images may have different levels of quality, which needs to be evaluated quantificationally. Nowadays, the mainstream full-reference (FR) metrics are effective to predict the quality of compressed images at coarse-grained levels (the bit rates differences of compressed images are obvious), however, they may perform poorly for fine-grained compressed images whose bit rates differences are quite subtle. Therefore, to better improve the Quality of Experience (QoE) and provide useful guidance for compression algorithms, we propose a full-reference image quality assessment (FR-IQA) method for compressed images of fine-grained levels. Specifically, the reference images and compressed images are first converted to $YCbCr$ color space. The gradient features are extracted from regions that are sensitive to compression artifacts. Then we employ the Log-Gabor transformation to further analyze the texture difference. Finally, the obtained features are fused into a quality score. The proposed method is validated on the fine-grained compression image quality assessment (FGIQA) database, which is especially constructed for assessing the quality of compressed images with close bit rates. The experimental results show that our metric outperforms mainstream FR-IQA metrics on the FGIQA database. We also test our method on other commonly used compression IQA databases and the results show that our method obtains competitive performance on the coarse-grained compression IQA databases as well.