Image Quality Assessment for Gaussian Blur using Siamese Network combined with ResNet-18
by Emrah Arslan, Omid Zare, Zeinab Mohseni, Mahdi Beigzadeh, Abel Abebe Bzuayene, Ali Abbaszadeh Sori, Javad Hassannataj Joloudari, Bulbula Kumeda Kussia
AI Insights
, Vol.1, No.2, 2025;
35 Views
This paper presents a novel Image Quality Assessment (IQA) framework, SNR (Siamese Network with ResNet-18), specifically designed for Gaussian blur detection. The approach leverages a Siamese network architecture combined with the ResNet-18 backbone to process image pairs—one blurred and one reference—to predict image quality based on their differences. The model effectively captures high-frequency features lost due to blur, such as edges and gradients. We conduct extensive experiments on the TID2013 dataset, showing that SNR achieves superior performance in blur-specific IQA tasks compared to other full-reference methods. Data augmentation techniques significantly improve model generalization, resulting in a test accuracy of 97.37% for ResNet-18. The proposed method demonstrates a strong correlation with human judgment and robust generalization across various image contents, with future work focusing on expanding its applicability to other distortions and optimizing computational efficiency.