Hybrid Perceptual Structural and Pixelwise Residual Loss Function based Image Super-Resolution

• Published in: 2024 3rd International Conference for Innovation in Technology (INOCON) pp. 1 - 6
• Main Authors: Singla, Shivam, Bohat, Vijay Kumar, Aggarwal, Manav, Mehta, Yatin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.03.2024
• Subjects: computer vision; image super-resolution; image-processing; machine learning; Noise reduction; Refining; Satellites; Superresolution; Surveillance; Technological innovation; Training
• DOI: 10.1109/INOCON60754.2024.10511465

Single Image Super-Resolution (SISR) aims to generate high-resolution (HR) images from the low-resolution (LR) input images. Deep learning techniques have significantly improved the HR outputs. These techniques primarily use the Mean Squared Error (MSE) Loss function for training purposes which considers pixels and doesn't take into account the structural information, textures, luminance, and contrast. These features play a crucial role in determining the perceptual quality of the image. Structural Similarity Index Metric (SSIM) can be manipulated to be used as a loss function that gives bright and well-structured HR images. However, this technique sometimes produces blurry and noisy HR outputs. To address this, a Hybrid Perceptual Structural and Pixel-wise Residual (HyPSPR) loss function approach has been introduced. This regularisation technique balances both factors. The Perceptual Structural (PS) Factor improves the structural information, textures, luminance, and contrast, while the Pixel-wise Residual (PR) Factor improves the encoded pixel-level information. A denoising and sharpening block in the model architecture has also been added to reduce the gaussian noise in the output.