RTSIQA: A database and method for real-world traffic scenes image quality assessment

Abstract

Traffic scene image quality assessment (IQA) is critical for intelligent transportation systems and autonomous driving applications. However, existing IQA methods are primarily designed for general real-world scenes and struggle to adapt to the structured elements and statistical characteristics unique to traffic scenes. Moreover, these methods overlook the distinct assessment needs arising from the spatially imbalanced perceptual importance in traffic scenes: some small regions (e.g., vehicles, pedestrians, traffic signals) are vital for driving safety, whereas some large regions (e.g., sky), despite their spatial dominance, are less critical. In addition, different traffic objects exhibit distinct degradation patterns due to their unique physical properties and texture structures, rendering a global quality score insufficient to represent differences in quality among these elements. Furthermore, the lack of IQA databases specifically for real-world traffic scenes has constrained further research development. To address these challenges, we construct a new real-world traffic scene IQA database providing both whole image quality scores and per-category quality scores for traffic object categories. Furthermore, we develop an adaptive multi-branch no-reference IQA network based on a dual-network architecture. This network extracts multi-scale features through pre-trained Swin Transformer combined with a semantic structure compensation module to enhance local structure modeling capability. It introduces a multi-branch assessment module utilizing object detection to identify traffic object location and category, achieving differentiated quality assessment for various traffic object categories. Experimental results show that the proposed method effectively outputs image quality for different objects within the same image on our constructed database and performs excellently on multiple general IQA databases.