Image Quality Enhancement using Deep Convolutional Network

People no longer tolerate images with low resolution or poor quality (color and contrast). However, certain criteria ensure the accessibility of such images in both professional and personal lives. Security monitoring cameras create low-resolution images to optimize limited bandwidth and storage space. While mobile phones are more prevalent in people's daily lives than professional-grade cameras, their photography quality remains inferior due to hardware gaps in sensors and chips. Lighting conditions and photographic expertise also limit the visual quality of captured images. Nowadays people show an interest in image enhancement to improve image quality. However, enhancing an image from a low-resolution source with poor color or contrast quality remains a challenging task. In this research, three different Convolutional Neural Networks (CNNs): Deeply-Recursive Convolutional Network (DRCN), Super-Resolution Convolutional Neural Network (SRCNN), and Super-Resolution Residual Network (SRResNet) are employed for image enhancement. For the research, image samples were collected from the DIV2K dataset. Upon comparing all three models, SRResNet performed well in image enhancement with a minimal Mean Squared Error (MSE) of 22.05, Root Mean Squared Error (RMSE) of 4.69, and higher Peak Signal-to-Noise Ratio (PSNR) of 37.21, along with a Structural Similarity Index (SSIM) of 0.976.