Research Projects

Underwater image enhancement is an important research topic in underwater computer vision techniques such as ocean navigation, sea-life monitoring, hydrothermal vent analysis etc. The underwater images are distorted due to various factors affecting the image formation process. Color distortion, blurriness and haziness are some of the common degradations that affect the underwater imagery. Image enhancement under the presence of distortions is a tedious process, as the distortions tend to enhance with the data. Therefore, appropriately choosing the parameters for the enhancement model is a crucial step in the enhancement process. Furthermore, some of the distortions cannot be modeled mathematically using a linear transform or an operator so as to perform the restoration as a simple matrix-inversion process. The distortion due to environmental conditions (under the water) manifest as random noise interventions making the enhancement process more challenging. Moreover, the images captured under severe (low) lighting conditions cause contrast degradation in captured images. Some portions of the images are under-exposed whereas some other portions get overexposed due to the lighting conditions under which the images are captured. Therefore, the uneven contrast aberration causes the enhancement and restoration, a real tedious step in underwater computer vision applications. The resolutions under the study should be able to handle various distortions simultaneously while restoring and enhancing the images. The non-linear modelling of the distortion problem demands a thorough theoretical analysis before experimentally verifying and deploying the same in the preprocessing step. Multiple distortions along with random noise interventions make the problem more challenging to solve using the naive restoration and enhancement methods available in the literature. Therefore, a deep learning based approach is being studied for handling these distortions in an efficient manner. A deep image prior model works with a single input image thus alleviating the need for a large number of images for training the model. A retinex based approach can handle color distortions and contrast degradations to a considerable extent. The retinex framework reduces the risk of over-enhancement of noisy features by appropriately decomposing the input data into disjoint components in terms of luminance and reflectance. Furthermore, an attention-based framework improves the detail-preservation capability of the model. The proposed strategy combines all these features into a single framework to handle multiple distortions (such as contrast degradation, intensity inhomogeneity, noise interventions, haziness, color distortions, dispersion artifacts, and blurring artifacts) simultaneously. The model needs to be verified theoretically to ensure the stability and convergence before experimentally analyzing the efficiency and deploying the same in underwater computer vision applications.