Weighted Robinson Compass Gradient and Charbonnier Penalty Function as a Loss Function
U. Rachna, Dhruv Shindhe S, S. N. Omkar · 2022
Loss function quantifies how well a machine learning model is performing, it computes the difference between the ground truth and the reconstructed outcomes. In image processing tasks such as image dehazing, image deblurring and image denoising the most popular loss functions used are Mean Squared error(MSE) and perceptual loss. MSE is less computationally expensive but fails to recreate precise boundaries whereas perceptual loss is more computationally expensive as they use feature vectors from a pre-trained model like VGG-net. In smaller on the edge systems like autonomous navigation systems, one of the most critical tasks is relying on visual cues, the take in the video feed from the environment and make decision on it after a few prepossessing steps to eliminate things like physical barriers, blurr, and smoke that clutter a robot’s environment are one of the several factors that affect autonomous navigation. We need systems that aren’t computationally expensive but also have a good reconstruction. For this we propose a loss function that incorporates gradients along with MSE for faster and better reconstruction.