A method for synthesizing complex weather conditions in images to improve the performance of computer vision models

Abstract

In this paper, we investigate and implement the data augmentation method to improve the performance of computer vision models when working in difficult visibility conditions. We analyzed possible approaches to estimating the three-dimensional structure of a scene and considered the principle of operation of a neural network model for depth estimation based on a single image. An algorithm for synthesizing weather conditions related to atmospheric scattering phenomenon using depth map and dark channel method was developed. An experiment is conducted to apply the considered augmentation method for the pre-training of the object recognition model. Performance verification on a set of real data showed that the presence of synthetic data in the training sample improves the recognition accuracy.