DOI: 10.14704/nq.2022.20.11.NQ66212

Underwater Image Enhancement Using D-Cnn

Suman Kumar Swarnkar, Gurpreet Singh Chhabra, Abhishek Guru, Bhawna Janghel, Prashant Kumar Tamrakar, Upasana Sinha

Abstract

Due to light dispersion and absorption, underwater photographs often exhibit colour distortion and reduced visibility. Existing techniques make use of a variety of presumptions and constraints to arrive at plausible improvements for underwater picture enhancement. The accepted assumptions may not hold true for some scenarios, which is a typical shortcoming of these methodologies. This research offers an end-to-end architecture for underwater image enhancement to solve this issue and introduces D-CNN, a CNN-based network. Color correction and haze removal are the two exercises used to train the D-CNN. With this dual training method, it is possible to concurrently learn a powerful feature representation for both tasks. The suggested learning framework considerably enhances the convergence speed and accuracy by using a pixel disruptive method to better extract the intrinsic characteristics in local patches. We create 200000 training photos based on the physical underwater imaging model to manage the training of D-CNN. Benchmark underwater photographs were used in experiments to compare D-CNN performance to those of other approaches.

Keywords

underwater image enhancement, deep CNN, color correction, haze removal

Full Text

References