基于拉普拉斯金字塔的图像压缩与重构研究

Image Compression and Reconstruction Based on Laplacian Pyramid

目的研究利用深度学习辅以拉普拉斯金字塔来完成图像压缩与重构。方法利用卷积神经网络提取图像的主要特征,利用双三线性插值法来减少特征尺寸,使用拉普拉斯金字塔来构建分层体系,从而逐步地减少图像大小以达到压缩的目的。在重构端上,对此系统则进行卷积操作,并采用上采样过程,进行图像的恢复重构过程,得到重构图。结果采用来自法国贝尔实验室的set 5与set 14数据集进行验证,使用2层金字塔即在16倍的高倍率压缩下进行实验结果验证,结果表明在主观评价上使用深度学习的方法在清晰度和还原度上要优于PCA,DCT和SVD,同时在客观评价上文中方法取得了标准差(52.73)与信息熵(7.44)的最好结果,高于PCA的49.70与7.38。SVD变换法与DCT变换法,在标准差上只有48.69和49.02,远不如文中方法,同时图片的信息熵只有7.34与7.35,低于文中的7.44。结论利用拉普拉斯金字塔结构来设计卷积神经网络结构来完成图像压缩与重构取得了不错的效果。

The paper aims to complete image compression and reconstruction through deep learning supplemented by Laplacian pyramid.Main features of the image were extracted with the convolutional neural network.The feature size was reduced by the bicubic linear interpolation.The hierarchy system was constructed by Laplacian pyramid to gradually reduce the image size and achieve image compression.On the reconstruction end,the corresponding convolution and up-sampling process was performed on the system;and the image reconstruction and reconstruction process was performed to obtain a reconstructed graph.Set 5 and set 14 from Bell Laboratories in France were used for verification.The experimental results were verified by the two-layer pyramid,which meant that the experimental results were verified at the 16 times of high-rate compression.The results showed that the method of deep learning was superior to PCA,DCT and SVD in terms of clarity and reduction in subjective evaluation,and the best results of standard deviation(52.73)and information entropy(7.44)were obtained in objective evaluation,which were higher than 49.70 and 7.38 of PCA.The standard deviations of SVD transform and DCT transform were only 48.69 and 49.02,which were far worse than the methods in this paper.Meanwhile,the information entropy of images was only 7.34 and 7.35,which was lower than 7.44 in this paper.Design convolutional neural network structure by Laplacian pyramid to complete image compression and reconstruction achieves good results.