基于细节增强的多能DR图像融合网络

Detail Enhancement-Based Fusion Network for Multi-Energy Digital Radiography Images

针对单能X射线无法对复杂工件各部位同时曝光成像的问题,提出一种基于细节增强的多能DR融合网络-双编码巢式连接融合网络。该网络以Inception模块作为基础卷积层,在双编码器的辅支路设计了可训练的LOG卷积模块,用来提取多尺度边缘特征,并将其补充到主支路以增强全局特征。在训练阶段,提出一种基于图像块的局部能量一致性损失函数,以减少输入、输出的局部性误差。融合时,采用通道和空间注意力机制作为融合策略,对双编码提取的多尺度增强特征进行融合,并将融合后的多尺度特征输入嵌套连接的解码器进行重构。结果表明,该融合网络具有细节增强效果,能够完整清晰地再现复杂工件的内部结构及缺陷。

To solve the problem that X-ray with single energy could not simultaneously expose each part of a complex workpiece,a multi-energy digital radiography(DR)fusion network based on detail enhancement,namely dual-encoder nest connection-based fusion network,is proposed.In this network,the inception module is used as the basic convolution layer and a trainable LOG(Laplacian of Gaussian)convolution module is designed in the auxiliary branch of the dual-encoder to extract multi-scale edge features and add them to the main branch to enhance the global features.In the training stage,a local energy consistency loss function based on image block is proposed to reduce the local errors of input and output.In the fusion process,channel and spatial attention mechanisms are used as the fusion strategy to fuse the multi-scale enhanced features extracted from dual-encoder,and the fused multi-scale features are input into the nest connected decoder for reconstruction.Experimental results show that the proposed fusion network has the effect of detail enhancement and can reproduce the internal structure and defects of complex workpiece completely and clearly.