Improved single image dehazing methods for resource-constrained platforms

Abstract Image dehazing is an increasingly widespread approach to address the degradation of images natural environment by low-visibility weather, dust and other phenomena. Advances in autonomous systems platforms have increased need for low-complexity, high-performing techniques. However, while recent learning-based image approaches significantly performance, this has often been at expense complexity hence use prior-based persists, despite their lower performance. This paper addresses both these aspects focuses on single dehazing, most practical class A new Dark Channel Prior-based algorithm presented that improved atmospheric light estimation method a low-complexity morphological reconstruction. In addition, novel, lightweight end-to-end network proposed, avoids information loss significant computational effort eliminating pooling fully connected layers. Qualitative quantitative evaluations show our proposed algorithms are competitive with, or outperform, state-of-the-art techniques with complexity, demonstrating suitability resource-constrained platforms.