Research On Image Inverse Halftoning Method Based On Multi-scale Feature Fusion Of Gan Network

Halftoning is a technology for transforming the continuous-tone image into binary image based on the low-pass characteristics of human vision,which solves the problem of continuous-tone image output and display in a limited tone equipment.However,this image quantization process is usually accompanied by irreversible image quality degradation,such as pixel and color loss.In order to get better reprocessing results,it is necessary to convert the halftone image into a continuous-tone image using the inverse halftoning technique firstly,and then reprocess the digital image.The traditional inverse halftone algorithm usually uses the prior knowledge of halftone to model manually,as a result,the generalization ability of the model is insufficient,and the quality of the restored continuous-tone images need to be improved.With the emergence of techologies such as artificial intelligence and convolutional neural networks,it has brought a new chance for inverse halftone technology,but most of algorithms use mean square error(MSE)and other pixel-level losses to optimize,resulting in blurry images.In view of above problems,this paper has carried out the following research on inverse halftone technology:(1)To impove the quality of continuous-tone image generated by the convolutional neural network,a novel inverse halftoning method based on multi-scale discriminator is proposed,which can enhance the restoration of image details by introducing generative adversarial network to learn the distribution characteristics of data.The generator designs a content aggregation sub-network and a detail enhancement sub-network with reference to the idea of a multi-task network,and gradually generates an inverse halftone image from coarse to fine.In order to capture more texture characteristics of the continuous-tone image,a multi-scale discriminator is selected to judge the image from multiple scales to further improve the performance of the model.(2)On the basis of the image inverse halftone network of the multi-scale discriminator,the MSE loss of the edge feature is introduced,and it is combined with content loss and adversarial loss as the loss function of the generator to further improve the image quality.In terms of objective and subjective evaluation,the results show that the proposed method can improve the visual effect of inverse halftoning image,and it is far better than other algorithms in improving edge detail restoration,and generating smoother contour lines.(3)To solve the problem that halftone artifacts remain in part of the generated image,a novel inverse halftoning method is proposed by fusing multi-scale feature.Designing a multi-scale feature extraction module,to enrich the correlation between the current pixel and the surrounding pixels and solve the problem that single size convolution kernel can not effectively be used to mine the multi-scale features of the image.This module extracts features of halftone images in parallel by using multiple dilation rates of dilated convolution,so that the network can obtain richer content information at the input,finally merge the features through channel splicing technology.The introduction of dilated convolution can expand the receptive field without changing the image resolution and model parameters,which is beneficial to improve the recovery effect of the image.(4)Since the information extracted by dilated convolutions contributes differently to the generated inverse halftone image,this paper introduces a channel attention mechanism to adaptively weight and fuse features of different scales to highlight the channel specificity of multi-scale features.The experimental results show that the final inverse halftoning image not only has better detail in the edge of the image,but also has better smooth characteristics,and the adaptability of this method to multiple halftones is also improved.The visual effect of the generated image is more realistic.