Quality Enhancement Of Terahertz Images Based On Generative Adversarial Networks

Terahertz imaging is a non-invasive imaging technique that has high application value in the fields of aerospace material testing and biomedical imaging,due to its low radiation and strong penetration properties.However,the original terahertz images suffer from a large amount of noise and low resolution due to system and imaging algorithm limitations,which makes it difficult to provide effective assistance for various detections.Therefore,how to improve the quality of terahertz images has become an emerging research hotspot.This thesis focuses on the image problems in the imaging results of the 110 GHz terahertz tomography imaging system,and proposes and designs a terahertz image processing method based on generative adversarial networks to improve the quality of terahertz images.The main research contents and contributions of this project are as follows:1.This thesis proposes a wavelet adaptive threshold denoising algorithm,which analyzes the wavelet coefficients obtained after image decomposition and sets the wavelet denoising threshold hierarchically to achieve denoising and reconstruction.It has been verified that this algorithm can effectively filter out image noise and stripe artifacts.Compared with the hard threshold wavelet denoising algorithm,the algorithm improves the peak signal-to-noise ratio of the Lena image results by 1.2d B and the structural similarity by 0.02.Further validation on terahertz images shows that the peak signal-to-noise ratio and structural similarity of the processed images are improved by about 3d B and 0.2,respectively.2.This thesis proposes an image super-resolution generative adversarial network based on a multi-scale feature extraction module.The network extracts deep features of the image and further uses generative adversarial network to achieve super-resolution reconstruction of the image.It has been verified that the algorithm achieves a peak signal-to-noise ratio of 36.85 and a structural similarity of 0.9457 for the 2x superresolution results of the Set5 test set images,which is an improvement compared to other algorithms.Validation on terahertz images shows that the peak signal-to-noise ratio and structural similarity of the processed terahertz tomography imaging results are improved by about 4.5d B and 0.33,respectively.3.This thesis proposes an image super-resolution generative adversarial network based on an improved attention mechanism.The network effectively captures and encodes key points in the image through the spatial attention mechanism,and adaptively adjusts the weights of the channels during training through the channel attention mechanism,making the network focus more on the useful features and suppressing the training of irrelevant information.It has been verified that the performance of the network is further improved while reducing half of the parameters.The algorithm achieves a peak signal-to-noise ratio of 36.95 and a structural similarity of 0.9491 for the 2x super-resolution of Set5 images.Validation on terahertz images shows that the peak signal-to-noise ratio and structural similarity of the processed results are improved by about 5d B and 0.4,respectively.The above results demonstrate that the terahertz image super-resolution reconstruction algorithm proposed in this thesis can effectively improve the quality of terahertz images and further assist the application of terahertz imaging in industrial and medical fields in terms of imaging detection.