End-to-End Blind Image Quality Assessment Using Deep Neural Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Blind Image Quality Assessment (BIQA) is crucial for evaluating image fidelity without reference images.
• Existing deep learning models for BIQA often face challenges with model complexity and training efficiency.
• Developing robust and efficient BIQA methods is essential for various image-related applications.

Purpose of the Study:

• To propose a novel multi-task end-to-end optimized deep neural network (MEON) for effective BIQA.
• To introduce a two-step training strategy that leverages readily available distortion identification data.
• To explore the use of biologically inspired Generalized Divisive Normalization (GDN) as an activation function for improved efficiency.

Main Methods:

• MEON utilizes two sub-networks (distortion identification and quality prediction) sharing early layers.
• A two-step training approach is employed: first, training the distortion identification sub-network, then the quality prediction sub-network.
• Generalized Divisive Normalization (GDN) is used as the activation function, replacing Rectified Linear Units (ReLU).

Main Results:

• The proposed MEON achieved state-of-the-art performance on four public BIQA benchmarks.
• GDN activation function demonstrated effectiveness in reducing model parameters and layers with comparable performance.
• MEON showed strong competitiveness against existing state-of-the-art BIQA models.

Conclusions:

• MEON offers a computationally efficient and high-performing solution for blind image quality assessment.
• The two-step training and GDN activation contribute to MEON's modest complexity and superior performance.
• MEON represents a significant advancement in the field of image quality evaluation.