Enhancement Of Image Datasets Based On Generating Adversarial Networks

Autonomous driving has become the future trend of intelligent vehicle development,and autonomous driving needs object detection technology to adapt to the changing environment around the vehicle.Object detection technology based on deep learning needs massive datasets for training to adapt to complex detection environment.Although Generation Adversarial Network can enhance the virtual dataset for target detection by image conversion,there are still some problems such as the lack of vehicle shadow data and multi-label natural environment data.Based on the improvement of the generated adversarial network,this paper studies the enhancement of image data.The main work is as follows:Firstly,Cyclic Consistent Adversarial network(Cycle GAN)and Multi-domain Unified Generative Adversarial Network(Star GAN)are reproduced experimentally,and the existing problems of both are explained.Cycle GAN can convert two image domains into each other,but without the supervision of local shadows,the shadow conversion effect is not good.Star GAN can use a single generator to transform multiple image attributes at the same time,but in the new scene of diversified transformation of natural environment,there is a lack of suitable dataset,which makes it impossible to conduct adaptive research.Secondly,based on Cycle GAN,a network framework V-Shadow GAN for multi-direction vehicle shadow conversion is proposed.In this framework,a shadow extraction and discrimination module is constructed to extract the shadow mask and make the authenticity discrimination,which guides the generation of more accurate vehicle shadows,and the channel attention mechanism SENet is added to the generator network to improve the output image quality.The unpaired vehicle shadow image dataset is made,including the vehicle shadow data in different directions in the morning,noon and afternoon.Experimental results show that the shadow images of vehicles in different directions in road scenes are successfully generated based on V-Shadow GAN.Subsequently,V-Shadow GAN algorithm is verified by comprehensive experiment.Compared with the standard image conversion algorithm,the Brenner and NRSS scores of V-Shadow GAN in shadow removal are 1055.4653 and 0.6641,respectively.The results were higher than those of Cycle GAN(474.1409 and 0.3893)and Mask-Shadow GAN(969.3836 and 0.6383).The Brenner and NRSS scores in shadow generation are 960.7025 and 0.6725,respectively,which are higher than 412.2053 and 0.3982 of Cycle GAN and 873.9264 and0.6618 of Mask-Shadow GAN,which verifies the effectiveness of V-Shadow GAN.The ablation experiment is carried out for each part of the algorithm.The ablation results show that the Brenner and NRSS scores of the whole V-Shadow GAN framework are higher than those of the individual parts,which proves the effectiveness of the integrated parts.Through the comparison experiment of different attention mechanism modules,it is proved that the selected channel attention mechanism module SENet has better effect.Finally,the multi-label and multi-scene natural environment dataset is researched and made,and Star GAN is improved to generate multi-scene natural environment data.A dataset of natural environment images with corresponding multi-scene labels is produced,in which the images have eight categories of labels under three scenes,including time,season and weather.The channel attention mechanism is added to Star GAN generator,and the upper sampling layer and convolution layer are replaced in U-Net structure to restore the image,which improves the effect of image generation.The experimental results show that the natural environment images of different scenes are successfully generated by using multi-label and multi-scene natural environment dataset.