|
Image completion is an important research direction in computer vision and has broad application prospects. Deep learning based image completion methods are generally based on three technologies, namely the autoencoder-based method, the generation of adversarial networks based method and the recurrent network based method. However, the output results of most methods are very single, for each masked Image input can only generate one completion result. Because the probability space corresponding to the possible results of each defective image is very large, in order to obtain the diversity of the completion results, this paper proposes an image completion method based on label differentiation, called LD-PICNet (Label Differentiation PICNet) This method can not only generate a complete image with clear and good semantic information, but also actively edit the tags on the generated results to maximize the diversity of the output. Specifically, this paper introduces an auxiliary classifier similar to ACGAN, which uses a single label of the image ground truth to reconstruct the image, and uses the label to actively increase the difference of the latent vectors during image completion to achieve variability of output. In addition, this paper also introduces a depth-weighted loss function through information entropy. The deeper the position of the image masked area, the lower the weight is given to further enhance the ability of the model to diversify the output. In order to evaluate the ability of LD-PICNet, this paper conducted experiments on 4 different data sets, and tested the model's ability to complete different types of targets, namely face (CelebA), architecture (Paris), landscape (Place2) ,and ordinary pictures (ImageNet). The results show that this method has the ability to generate diversified results, and it has higher clarity than the more advanced methods. |
|
Keywords:Image processing; image completion; generation of adversarial networks; label differentiation |
|