A Multi-Level Patch Dataset for JPEG Image Quality Assessment by Absolute Binary Decision

Abstract

Image quality assessment (IQA) plays a fundamental role in evaluating image processing. Currently, JPEG AIC specifies the IQA methods, dividing them into three levels: AIC-1, 2, and 3. AIC-1 measures the quality from low to high, AIC-2 focuses on the threshold for visual losslessness, and AIC-3 measures the range between 1 and 2. AIC-3 requires complex processing and many comparisons, such as using boosted triplets to obtain highly accurate JNDs and then using those JNDs to create scale scores, or generating many combinations of triplets. In this study, we revisit the definition and propose a method for measuring the target band of AIC-3 by mixing the measurement methods of AIC-1 and AIC-2 and adjusting the sensitivity. This method presents the pristine and degraded images and asks whether they are the same or not. We called this absolute binary decision (ABD), referring to ACR in AIC-1. We constructed a JPEG-specific IQA dataset using ABD from distorted images that were progressively patched to relate the patches to the IQA of the entire images. As this was a new experiment, it was first conducted under laboratory control to ensure reliability. The experimental results showed that ABD could measure the QP40-90 range. In addition, it was found that patching differs from the entire image case. While patching draws attention to places that people do not usually pay attention to, usual image presentation concentrates attention through semantic guidance, suggesting the possibility that pseudo-attention patching is being performed on characteristic locations.