目前用于生成X射线图像的方法中存在主体过拟合和背景欠拟合等问题,针对此类问题,基于去噪扩散概率模型DDPM(Denoising Diffusion Probability Model)提出了一种新型图像生成模型MDDPM(Masked DDPM),设计一种无监督图像分割方法对X射线图像进行分割,将分割后得到的二值图像作为遮罩加权到损失函数,增强扩散模型;设计一种含有增强型SE注意力块的卷积块ESE Block(Enhanced Squeeze-and-Excita-tion Block),结合注意力机制和上、下采样模块等搭建U-Net结构的神经网络,进一步提高网络的学习、表征和泛化能力.使用MDDPM在OPIXray数据集上验证了对X射线违禁品图像进行增广的可行性,针对五个类别的违禁品,实验结果表明,相比于DDPM,MDDPM的生成图像质量分布差异指标FID分别提升了18.3%、24.82%、32.85%、29.12%和33.62%.将使用本模型生成的图像与原始图像进行混合,与只使用原始图像进行图像分类实验相比,分类精确度提高了3.2%,此结果表明,生成的图像不仅保留了原始数据的特征,而且提高了数据高维特征的多样性.
Mask-Enhanced Diffusion Model For X-Ray Image Generation
Current methods used to generate X-ray images have problems such as subject overfitting and background under-fitting.To address the above problems,the paper proposes a new image generation model Masked DDPM(MDDPM)based on the Denoising Diffusion Probability Model(DDPM),designs an unsupervised image segmentation method to segment X-ray images,and uses the binary image obtained after segmentation as a mask to weight the loss function to enhance the diffusion model;in addi-tion,a convolution block ESE Block(Enhanced Squeeze-and-Excitation Block)with enhanced SE attention block is designed,com-bining attention mechanism and up-and-down sampling modules to build a U-Net structured neural network to further improve the learning,representation and generalization ability of the network.The feasibility of augmenting X-ray contraband images was veri-fied using MDDPM on the OPIXray data set.For five categories of contraband,the experimental results show that compared with DDPM,the generated image quality distribution difference index FID of MDDPM is improved respectively 18.3%,24.82%,32.85%,29.12%and33.62%.The images generated by using this model are mixed with the original images.Compared with u-sing only the original images for image classification experiments,the classification accuracy is increased by 3.2%.This result shows that the generated images not only retain the characteristics of the original data,but also improve the diversity of high-dimen-sional features of data.
万方数据
维普
