Mask-guided Modal Conversion Algorithm Based On Generative Adversarial Networks

In the field of computer vision based on deep learning techniques,the task of image to image translation is an important area of research.The task requires the model to learn a mapping relationship between an input image to an output image,thus enabling the translation between images of different modalities.In this thesis,we propose a mask-guided image modal translation algorithm based on generative adversarial networks to address the situation that existing modal translation algorithms cannot maintain a strictly consistent topology of the input image to the output image and cannot accommodate the problem of grey scale inconsistency between different individuals of the same modality.The algorithm allows the network to predict the results while keeping the input and output images strictly consistent in topology.In addition,the adaptive coding module proposed in this thesis allows the network to guide the modal conversion according to a specific reference image,so that the converted image can have both the modal properties of the target neighbourhood and the appearance grey scale properties of the reference image.The research work in this thesis is as follows.1)In this thesis,a mask-guided image modal conversion algorithm based on a generative adversarial network is designed for a specific application scenario.The algorithm generates a greyscale adjustment mask through a generator with an encodingdecoding structure and uses a residual structure to sum this mask directly with the input image to obtain the transformed resultant image,and introduces an edge constraint loss to further enhance the control of the generated image topology.Experiments on different MRI datasets of brain images show that the algorithm proposed in this thesis can achieve not only the optimum in image quality metrics,but also the leading in image structure similarity metrics.2)For the problem of grey scale inconsistency,an adaptive style mask-guided image modality conversion algorithm based on reference maps is proposed.The algorithm enables the model to extract the overall features of the target neighbourhood while embedding the grey scale features of the reference image.Experiments show that the algorithm proposed in this thesis can not only solve the grey-scale inconsistency problem,but also achieve optimality in other image quality evaluation metrics.Experiments at the application level of modal transformation show that the proposed algorithm as a data enhancement method can improve the accuracy of image registration tasks,which further proves that the proposed image modal conversion algorithm has high practical value.