Research Of The Cross-domain Image Understanding Based On Generative Adversarial Neural Networks

Nowadays,deep learning is known by more and more people,and its application is more and more extensive,most of model in the field of computer vision are deep learning networks.The key to the effectiveness of deep learning network lies in training,the ultimate goal is to train a model,which can be used to make a series of predictions.Training deep learning network requires a lot of datasets,the size and complexity of existing datasets have lagged behind the growth of model capacity,and data becomes a problem if you want to train and apply the new model.The most common way to obtain data is to manually obtain and annotate it,but this is a highly complex and costly method of obtaining data,the acquisition speed is slow when large amounts of data are required.Therefore,a scheme for converting easily obtained data into target data was proposed,which solved the problems of high cost,difficulty of acquisition,and slow acquisition speed.Then,the next step is to study how to better transform data.Cycle Generative Adversarial Network can achieve the above transformation function,that is synthesis of cross-domain images.The network is composed of two generators and two discriminators.However,there are many problems in the images generated by the generator after the training,the most prominent problem is that the visual content of the images cannot be well preserved in the cross-domain process,so that the cross-domain images cannot correspond to each object in the label.Therefore this paper proposes Label Preserving Generative Adversarial Network(LPGAN)combined with image filtering operation,can better keep the contents of the images after cross domain information and enhance object edge,so the generated image will corresponds to the label file,cross domain data is better to object detection network identification for the target domain data.LPGAN is mainly based on the Cycle Generative Adversarial Networks,which is combined with the semantic segmentation network.In LPGAN,the weight cross entropy function is added to the loss function,and the semantic segmentation loss results are used to constrain the changes of the content information in the pictures after the cross-domain.The scheme proposed in this paper is divided into two steps: 1)train LPGAN,and generate cross-domain images;2)filter the cross-domain image to make the image smoother and the edges clearer.According to the scheme experiment proposed in this paper,the object detection network,which is trained by our cross-domain images,can achieve higher mean average precision value on a test datasets,thus proving the effectiveness of the scheme in this paper.