Image Quality Assessment by Integration of Low-level & High-Level Features: Threshold Similarity Index

The increase in the implementation of deep learning models in image processing has shifted its focus towards better-performing models within the ambit, directly proportional to the size of the dataset. A common approach to increasing the available data for training is via data augmentation. Additional data is created either by modifying gathered data or creating synthetic data based on the seed dataset. This article presents an approach to validating synthetic images generated by a Deep Convolutional Generative Adversarial Network (DCGAN) using full-reference image quality assessment techniques. We have presented an algorithm based on the Threshold Similarity Index(TSI) to validate synthetically generated images. The TSI becomes a special index. TSI calculation has been done by amalgamating high-level features like luminance, contrast, structure, and low-level features like edges and zero crossings of the correlated images. TSI becomes an integrated matrix of our algorithm by incorporating the features of SSIM and FSIM both. The developed algorithm has been verified by generating synthetic images using the DCGAN model, which generated 35 images. The TSIFSIM and TSISSIM of this model were calculated to be 0.637 and 0.127 respectively. The proposed algorithm validated 85,7% of synthetically generated images to be included in the seed dataset.