Norm-in-Norm Loss with Faster Convergence and Better Performance for Image Quality Assessment

Abstract: Currently, most image quality assessment (IQA) models are supervised by the MAE or MSE loss with empirically slow convergence. It is well-known that normalization can facilitate fast convergence. Therefore, we explore normalization in the design of loss functions for IQA. Specifically, we first normalize the predicted quality scores and the corresponding subjective quality scores. Then, the loss is defined based on the norm of the differences between these normalized values. The resulting "Norm-in-Norm'' loss encourages the IQA model to make linear predictions with respect to subjective quality scores. After training, the least squares regression is applied to determine the linear mapping from the predicted quality to the subjective quality. It is shown that the new loss is closely connected with two common IQA performance criteria (PLCC and RMSE). Through theoretical analysis, it is proved that the embedded normalization makes the gradients of the loss function more stable and more predictable, which is conducive to the faster convergence of the IQA model. Furthermore, to experimentally verify the effectiveness of the proposed loss, it is applied to solve a challenging problem: quality assessment of in-the-wild images. Experiments on two relevant datasets (KonIQ-10k and CLIVE) show that, compared to MAE or MSE loss, the new loss enables the IQA model to converge about 10 times faster and the final model achieves better performance. The proposed model also achieves state-of-the-art prediction performance on this challenging problem. For reproducible scientific research, our code is publicly available at https://github.com/lidq92/LinearityIQA.

• The paper introduces the Norm-in-Norm loss function, which normalizes predicted and subjective scores to accelerate convergence and improve accuracy in image quality assessment models.
• Theoretical analysis shows this loss stabilizes gradients, while experiments demonstrate significantly faster convergence and better performance on standard image quality assessment datasets.
• This novel loss offers practical benefits for training efficiency with large datasets and theoretically opens new possibilities for integrating normalization into loss function design for various machine learning tasks.

Overview of "Norm-in-Norm Loss with Faster Convergence and Better Performance for Image Quality Assessment"

The paper presents a novel approach to image quality assessment (IQA) by introducing the "Norm-in-Norm" loss, which aims to address the slow convergence typically associated with models supervised by mean absolute error (MAE) or mean square error (MSE) loss. The authors propose leveraging normalization as a mechanism to enhance convergence speed and performance in IQA models.

Key Contributions

1. Introduction of Norm-in-Norm Loss: The study introduces a loss function that incorporates normalization of both predicted and subjective quality scores. The idea is to normalize scores by subtracting their mean and dividing by their norm before computing the loss. The resulting "Norm-in-Norm" loss is constructed as the $L^p$ norm of the differences between these normalized values.
2. Theoretical Insights: A connection is established between the proposed loss and common IQA performance metrics, specifically Pearson’s Linear Correlation Coefficient (PLCC) and Root Mean Square Error (RMSE). A variant of the loss is shown to align with RMSE, indicating that it accurately reflects objective quality measures.
3. Improved Convergence: The paper provides theoretical justification that the introduced normalization stabilizes the gradient dynamics, resulting in faster convergence of the IQA model. Empirical evidence supports this, showing convergence approximately ten times faster than when using MAE or MSE losses.
4. Empirical Validation: Through experiments on real-world IQA problems, particularly the quality assessment of in-the-wild images, the proposed model demonstrates better performance in terms of both prediction accuracy and convergence speed on datasets such as KonIQ-10k and CLIVE.

Implications and Future Work

The introduction of the Norm-in-Norm loss function has significant implications for the design of loss functions in regression-based tasks within IQA and potentially other computer vision domains. By promoting faster convergence and improved model stability, this approach can reduce computational costs and enhance the robustness of deep learning models.

Practically, the ability to converge models rapidly while maintaining high accuracy is crucial given the increasing size of datasets in the deep learning era. This contribution could influence the development of scalable IQA models fit for deployment in real-world scenarios where quick iteration and deployment cycles are essential.

Theoretically, the study opens new avenues for integrating normalization techniques within loss functions, potentially leading to novel loss designs that could benefit a broader array of machine learning tasks. Future research could explore the adaptability of the proposed loss across various data types and how its hyperparameters could be optimized for specific contexts.

Conclusion

The paper provides a significant contribution to the field of image quality assessment through the development of a novel loss function that effectively integrates normalization for enhanced model convergence and performance. This work demonstrates both theoretical depth and practical applicability, laying groundwork for further advancements in efficient machine learning model training techniques.