Evaluating Image Quality Metrics as Loss Functions for Image Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Automated image quality assessment (IQA) metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Metric (SSIM), are widely used due to the challenges of human evaluation.
• Discrepancies between automated metrics and human judgment necessitate the development of more accurate IQA methods.
• Traditional IQA metrics have been limited to post-hoc quality assessment, not integrated into neural network training.

Purpose of the Study:

• To evaluate the effectiveness of recent IQA metrics as loss functions for training neural networks.
• To compare the performance of IQA-optimized neural networks against those trained with standard loss functions.
• To investigate the potential of direct optimization for desired IQA metrics in high-level vision tasks.

Main Methods:

• Ten recent image quality assessment (IQA) metrics were evaluated.
• These IQA metrics were implemented as loss functions within two dehazing neural networks.
• The performance of networks trained with IQA loss functions was compared to those trained with conventional loss functions.

Main Results:

• Using advanced IQA metrics as training objectives led to broad improvements in neural network performance.
• The study demonstrated the efficacy of integrating IQA metrics directly into the training process.
• Optimizing for specific IQA metrics showed superior results compared to standard loss functions.

Conclusions:

• Recent IQA metrics are effective as loss functions for training image restoration networks.
• Directly optimizing neural networks for image quality assessment metrics can significantly enhance performance.
• This approach offers a promising direction for improving high-level vision tasks through better image quality optimization.