Reconocimiento del clima en imágenes utilizando aprendizaje profundo autosupervisado.

Abstract

Automatic weather recognition from digital images has important applications, such as: land and air traffic control, autonomous vehicles, road safety warnings, crop control, among others. Building deep learning models for weather recognition is a challenge for computer vision when it comes to extracting the deterministic characteristics of each type of weather. Therefore, several related works have proposed models based on binary classifiers, which allow mapping between positive and negative classes. These works have reported as main difficulties that the classes are not completely exclusive and the need for thousands or millions of labeled images for each class required for training deep learning models. To face several of the difficulties identified, in this research paper we propose to apply the concepts of "self-supervised deep learning". Self-supervised learning follows the traditional phases of deep learning and adds the stages of pre-training, transfer learning, and fine tuning. The pretraining stages are the ones that allow to reduce the number of labeled images required, while the later stages allow to generalize and improve the model. The Residual Learning Network Architecture, RESNET, has been selected for this work based on exhaustive research and several practical tests that allow to identify the benefits of using this architecture in the field of weather recognition. The proposed method was tested and adjusted to achieve metrics of 93% in detection rates.