Research On Automatic Object Removal Based On GAN

Target detection and image completion are two important research directions of computer vision.In the field of object detection,the fine labeled image data is the key premise for the model to achieve better performance as the model supervision condition.However,image annotation needs a lot of labor costs,and It is difficult to detect objects from a large number of weakly labeled images on the Internet.In the field of image completion,most of the existing technologies with better results are accomplished by object-centered image,such as face completion or structured scene completion,but in the general scene level image,the results are relatively general.Therefore,the core of this paper is how to use the image with only image level label to locate and delete the objects in the general scene image,and then through the image completion network to complete the holes left by the deleted objects.The main research work and achievements are summarized as follows.First of all,based on the weak supervision deep detection network,this paper achieves a multiple instance detection algorithm with instance classifier refinement.The algorithm can find target detection through the image level label.The framework combines the multi instance detection network with the multi-stage instance classifier,and proposes an algorithm to optimize the instance classifier.The algorithm integrates the basic instance detection network and the multi-stage instance classifier optimization network into a single network,which can realize the end-to-end training.In this paper,based on the existing work,the performance of weak supervised target detection is improved by applying the above improvements.Secondly,this paper improves the image completion network based on DCGAN,and implements a novel image completion method based on WGAN,which can keep the local and global consistency of the image.By using the complete convolution neural network,the missing area of any shape can be filled to complete the image of any resolution.Using space attenuation reconstruction loss instead of general reconstruction loss is more effective in improving the quality of completion when a large area of missing area is completed.Two discriminators,global and local,are used for training.The global discriminator looks at the whole image to evaluate whether it is consistent,while the local discriminator only looks at the small area centered on the completion area to ensure that the generated patches are consistent in the local area.Based on the existing work,this paper introduces W-GAN loss function,attention mechanism and the idea of phased completion to overcome the shortcomings of the existing methods,and achieve the improvement of the completion effect.In this paper,a large number of experiments are carried out for the above two different research tasks,and a number of comparative experiments are constructed.The experimental results show that,in qualitative and quantitative comparison with other methods,this paper achieves the improvement of weak supervised target detection performance and image completion effect.Through the realization of the above two functions,we can realize the function of automatic object removal: input an image into the multi instance detection network optimized by the instance classifier to obtain the coordinate position information of the object to be removed,and then remove the target object to obtain the image to be reconstructed without the target object,and input it into the image completion network for reconstruction and repair,and finally Get a good image without the target object.