Underwater Image enhancement using residual networks

Ankita Naik (arnaik), Apurva Swarnakar (aswarnakar)

Abstract: Due to physical phenomena like refraction, absorption, and scattering of light by suspended particles in water, raw underwater images have low contrast, blurred details, and color distortion. Such degradation of images interferes with computer vision tasks like segmentation, object detection and classification. Thus, underwater image enhancement is carried out for tackling such phenomenons. In the realm of deep learning models majorly GANs and CNNs are used for the task. Through our study, we work on drastically improving a residual network (UResnet) which works at power with state-of-art GANs for image enhancement. Our proposed architecture works towards improving the quantitative image enhancement metric while reducing the model complexity and computation requirements. The same is carried out by experimenting with loss functions generally utilized for image enhancement tasks like super-resolution reconstruction and improving the residual network architecture using novel skip connections.