Ding Xiaxu, Liu Yi, Yan Hongxu, Zhang Pengcheng, Guo Niu, Gui Zhiguo
Appl Opt. 2023 Jul 10;62(20):5526-5537. doi: 10.1364/AO.493750.
X-ray images frequently have low contrast and lost edge features because of the complexity of objects, attenuation of reflected light, and scattering superposition of rays. Image features are frequently lost in traditional enhancement methods. In this paper, we use a ray scattering model to estimate coarsely clear images and an encoder-decoder network and multi-scale feature extraction module to add multi-scale and detail information to the images. To selectively emphasize useful features, a dual attention module and UnsharpMasking with learnable correction factors are used. The results of the experiments demonstrate that the method may significantly enhance the quality of x-ray images.
由于物体的复杂性、反射光的衰减以及射线的散射叠加，X射线图像经常具有低对比度和边缘特征丢失的问题。在传统的增强方法中，图像特征经常丢失。在本文中，我们使用射线散射模型粗略估计清晰图像，并使用编码器-解码器网络和多尺度特征提取模块为图像添加多尺度和细节信息。为了有选择地强调有用特征，我们使用了双注意力模块和带有可学习校正因子的非锐化掩蔽。实验结果表明，该方法可以显著提高X射线图像的质量。
