摘要
对于一个图像处理系统,包括图像的获取、处理、传输、接收、输出等环节,都会存在不同程度的噪声,使图像质量降低,影响后续的处理。Transformers神经网络结构在自然语言和高级视觉任务上表现出显著的性能提升。虽然Transformer网络减轻了卷积神经网络在感受野和注意力机制上的不足,但其计算复杂度随空间分辨率成二次方增长,因此无法应用于大多数涉及高分辨率图像的图像恢复任务。本文基于改进型的Transformer网络架构,通过改进的注意力机制以及像素重塑模块,有效地降低了Transformer网络的计算复杂度,使得模型能够支持更高分辨率的图像输入,为模型提供了更好的图像细节,使得模型在图像去噪上的达到了优于以上传统方法和基于卷积神经方法的效果。
For an image processing system, including image acquisition, processing, transmission, reception, and output, there are varying degrees of noise that degrade the image quality and affect the subsequent processing. Transformer-based neural network structure shows significant performance improvement on natural language and advanced computer vision tasks. Although Transformer networks alleviate the deficiencies of convolutional neural networks in perceptual field and attention mechanisms, their computational complexity grows quadratically with spatial resolution and thus cannot be applied to most image recovery tasks involving high-resolution images. In this paper, based on the improved Transformer network architecture, the computational complexity of the Transformer network is effectively reduced by the improved attention mechanism and the pixel reshaping module, which enables the model to support higher-resolution image inputs and provides better image details for the model, making the model achieve better image denoising than the above traditional methods and convolutional-based neural methods.
出处
《计算机科学与应用》
2022年第12期2763-2771,共9页
Computer Science and Application
