Abstract
Blind Image Quality Assessment (BIQA) is a fundamental task in computer vision, which however remains unresolved due to the complex distortion conditions and diversified image contents. To confront this challenge, we in this paper propose a novel BIQA pipeline based on the Transformer architecture, which achieves an efficient quality-aware feature representation with much fewer data. More specifically, we consider the traditional fine-tuning in BIQA as an interpretation of the pre-trained model. In this way, we further introduce a Transformer decoder to refine the perceptual information of the CLS token from different perspectives. This enables our model to establish the quality-aware feature manifold efficiently while attaining a strong generalization capability. Meanwhile, inspired by the subjective evaluation behaviors of human, we introduce a novel attention panel mechanism, which improves the model performance and reduces the prediction uncertainty simultaneously. The proposed BIQA method maintains a lightweight design with only one layer of the decoder, yet extensive experiments on eight standard BIQA datasets (both synthetic and authentic) demonstrate its superior performance to the state-of-the-art BIQA methods, i.e., achieving the SRCC values of 0.875 (vs. 0.859 in LIVEC) and 0.980 (vs. 0.969 in LIVE).
Introduction
Evaluating the quality of an image without a "pristine" reference—known as Blind Image Quality Assessment (BIQA)—is notoriously difficult. Images in the wild suffer from diverse distortions such as blur, noise, and compression, alongside varied content. Current methods based on Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs) usually follow a "pre-train then fine-tune" pipeline. However, a significant gap exists: pre-training focuses on semantic content (identifying objects), while IQA requires focusing on perceptual details (how the image looks). Bridging this gap typically requires massive amounts of human-labeled data, which is expensive and labor-intensive to collect. Furthermore, human judgment is inherently subjective, leading to high prediction uncertainty in AI models.
Methodology
To address these challenges, we propose DEIQT (Data-Efficient Image Quality Transformer), which moves beyond the standard Transformer encoder-only architecture.
Quality-Aware Decoder
Instead of simply feeding the encoder's "CLS" (classification) token into a regression head, DEIQT utilizes a Transformer decoder. This decoder uses the CLS token as a query to "re-examine" the spatial features from the encoder via cross-attention. This process refines abstract semantic features into specific quality-aware features, acting as an "interpreter" that pulls out the quality signal hidden within semantic noise.
Attention-Panel Mechanism
Inspired by the way a panel of human subjects evaluates an image, we introduce $L$ learnable "panel member" embeddings. Each member interacts with the image features to provide a unique perspective on quality. The model generates a quality score for each panel member, and the final quality score is the average of these individual scores. This multi-member voting reduces prediction uncertainty and improves stability.
Experimental Results
DEIQT was evaluated on eight standard BIQA benchmarks, covering both synthetic (LIVE, CSIQ, TID2013, KADID) and authentic distortions (LIVEC, KonIQ-10k, LIVEFB, SPAQ).
Data Efficiency
A striking insight from our experiments is the model's data efficiency. DEIQT reaches the performance of previous state-of-the-art (SOTA) models (such as TReS) using significantly less training data. As shown in the figure below, the model matches benchmarks using only 30% of the typical training set.
Performance and Generalization
DEIQT achieved top-tier results across all datasets, reaching an SRCC of 0.875 on LIVEC and 0.980 on LIVE. Beyond accuracy, the model demonstrates strong cross-dataset generalization; when trained on one dataset and tested on another without fine-tuning, it consistently outperformed competitors. Despite its performance, the design remains lightweight, utilizing only a single decoder layer and roughly 24 million parameters.
Conclusion
DEIQT demonstrates that a decoder-based architecture with an attention-panel mechanism can significantly improve both the accuracy and data efficiency of BIQA. By treating fine-tuning as an interpretation of pre-trained features, the model stabilizes training and captures diverse perceptual aspects of image quality. Future work may explore predicting the full distribution of human scores to better capture subjective uncertainty.
Summary updated