基于宽度学习的集成超分辨率重建方法

Integrated super-resolution reconstruction method based on broad-learning

不同的超分辨率重建算法重建得到的高分辨率图像都各有所长,为将这些重建图像中的有效信息集成到一幅图像中,做到优势互补,提高图像的质量,提出基于宽度学习的集成超分辨率重建方法。利用径向基神经网络的宽度学习模式训练一个预测模型,利用该模型对新输入数据进行预测,得到其对应的权值,重建高分辨率图像。对常用的近邻重建算法、稀疏重建算法以及稀疏近邻算法进行集成,在医学图像上进行实验,实验结果表明,相较集成的算法,该算法重建效果有较大程度提高,比近邻重建算法PSNR提高1.14dB,SSIM提高0.23,比稀疏重建算法PSNR提高1.26dB,SSIM提高0.05,比稀疏近邻重建算法PSNR提高1.01dB,SSIM提高0.03。

The high-resolution image reconstructed using different super-resolution reconstruction methods has strength and weakness.Integrated super-resolution reconstruction method based on broad-learning was proposed to integrate the useful imformation of the reconstructed images into a single image,combine advantages and improve the quality of the image.By calculating the optimal nonnegative weights corresponding to each patch,and making full use of the RBF neural network based on broad-learning,and keeping adding neurons number until meeting the need of precision,aprediction model was trained.The corresponding weights of the new input data were obtained through the prediction model.The high-resolution image was reconstructed.Common methods such as neighbor reconstruction algorithm,sparse reconstruction algorithm and sparse neighbor embedding reconstruction algorithm were integrated,and methods were implemented on medical image images.Experimental results show that the proposed algorithm can achieve better reconstruction effects compared to the integrated algorithms with improved PSNR of 1.14 dB and improved SSIM of 0.23 compared to neighbor embedding-based reconstruction algorithm,and improved PSNR of 1.26 dB and improved SSIM of 0.05 compared to the sparse representation algorithm,and improved PSNR of 1.01 dB and improved SSIM of 0.03 compared to sparse neighbor embedding reconstruction algorithm.