Research On Single-pixel Imaging Method Based On Generative Adversarial Network

Single-pixel imaging is a frontier research hotspot in the field of computational imaging.The technology modulates the light field of the scene,uses a single-pixel detector without spatial perception to collect information,and reconstructs the scene image with fewer sampling values.The band of single-pixel imaging is not limited to the visible light band,which makes it have unique advantages in the fields of remote sensing imaging,terahertz imaging,and autonomous driving perception.The traditional reconstruction algorithm of single-pixel imaging relies on compressed sensing technology,and the quality of compressed sensing reconstructed image is difficult to guarantee at extremely low sampling ratios.In order to obtain better single-pixel imaging quality and faster imaging speed at extremely low sampling ratios,this thesis studies the single-pixel imaging method based on deep learning generative adversarial networks.This thesis first designs a single-pixel imaging model based on conditional generative adversarial networks.The generator uses a residual network structure,the discriminator is a Patch GAN structure,and the model also includes a feature expansion module.The Walsh-Hadamard matrix is used as the measurement matrix to obtain the measured values.The measured values are the input of the model.The measured values after feature expansion are used as the model condition to guide the generation direction of the model.The objective function includes adversarial loss and regularization loss,and the training model learns the end-to-end reconstruction mapping relationship from the measured values to the image.Under the condition of low sampling ratios,the model can reconstruct images with better quality.Secondly,in order to further improve the integrity of the reconstructed image structure,this thesis designs a single-pixel imaging model of progressive generative adversarial network with attention mechanism.The generator uses a progressive learning strategy to learn gradually from low resolution to high resolution.A hybrid attention mechanism is added to the generator to improve the efficiency of obtaining effective information from the previous stage.The objective function consists of adversarial loss,regularization loss and structural similarity loss.The image reconstructed by the model has a significant improvement in the evaluation index and the subjective visual effect is more delicate.The structural integrity of the reconstructed image is further improved.Finally,the two methods proposed in this thesis are simulated on MNIST dataset and STL-10 dataset,and the experimental results are analyzed and compared.In addition,this thesis summarizes the research work and proposes areas for improvement.The practical application of single-pixel imaging based on deep learning generative adversarial networks is prospected.