Um Método para a Tradução de Objetos em Imagens Digitais através de Redes Generativas Adversariais com Consistência Cíclica

Abstract

Image translation methods based on deep learning allow transforming an image from a given context into another image associated with a different context. The method proposed in this work is a particular case of image translation, where only certain objects present in an image are translated. The central idea is to stamp a symbol, or a standard representation of a class of objects, on the original image. This new image is then processed by a set of neural networks that make up an architecture called CycleGAN, or cycle consistent adversarial generative networks. In the final image, the stamped symbol appears realistically, incorporating features of the original image, such as lighting and visual distortions. To evaluate the proposed method, an application of great importance for autonomous vehicle driving technology was chosen, namely, the recognition of road signs in images of urban scenes. Due to the large number of existing road sign classes, and the considerable imbalance of samples from different classes, the proposed method was used to generate new images, replacing the road sign symbols in the original images. In this way, the proposed method was used to balance the samples of a public database of road signs, in a data augmentation scheme for training a previously proposed object detection and recognition model, the YOLOv3 . Visual analysis of the synthesized images showed that they have good quality in the sense of looking realistic. In addition, the accuracy of the model increased from 0.41 mAP to 0.54 mAP, for a database with originally 900 images and 43 classes We believe that this result quantitatively proves the effectiveness of the proposed method and points out how it is possible to use image translation techniques to add knowledge in the training of an object recognition network.