Research On Core Image Compression Method Based On Deep Learning

As an important part of oil and gas engineering,cores play an important role in exploration and mining,and core images naturally become the most research-oriented resources.With the continuous iterative update of the camera hardware,the size of the image of the core image and the required storage space are also increasing day by day,while the capacity of the storage device is developing very slowly.In order to store and manage more and more core images,the development of the spearhead is straight.Refers to the field of core image compression and storage.The core image itself has complex texture features.The traditional core image compression method uses wavelet transform and discrete cosine-based transform to compress the core image.However,these methods will lose the image part information and lead to the loss of image detail texture.To this end,this paper proposes two core learning methods based on deep learning,which can increase the compression ratio and reconstruct the high-definition core image to improve the visual effect.The reconstructed image has both SSIM index and PSNR index.Higher value.The main research contents and innovations described in this paper are as follows:Firstly,study the classic image compression literature at home and abroad,including traditional image compression algorithms and models;image compression algorithms and models based on deep learning;and image high resolution reconstruction methods based on deep learning.These include three kinds of traditional lossless coding,arithmetic coding,Huffman coding,and run-length coding,JPEG,predictive coding,vectorization coding,fractal coding,neural network coding,and JPEG2000 six traditional lossy compression coding.Image compression and HD reconstruction methods based on deep learning include: SRCNN model,FSRCNN model,ESPCN model,VDSR model,DRCN model,RED model,DRRN model,SRDenseNet model,EDSR model,SRGAN model,EnhanceNet model,IDN model,LapSRN Model,MemNe model,DBPN model,RDN model,RCAN model,MSRN model,CARN model,ZSSR model,SFTGAN model,SRFeat model,SERGAN model.In Chapter 2,the JPEG and JPEG2000 compression methods are introduced.Then the existing core image compression theory and algorithm are studied.The core image compression algorithm based on block compression sensing and the wavelet core image compression algorithm based on texture analysis are introduced in detail.EBCOT core image compression algorithm based on wavelet packet.Finally,the chapter introduces three common indicators to measure the quality of reconstructed images,including mean square error,peak signal to noise ratio,and SSIM.Secondly,after studying the theoretical knowledge of JPEG technology and convolutional neural network,the combination of high compression rate of JPEG and high feature extraction,high image reductiveness and high processing autonomy of convolutional neural network is proposed.Two improved core image compression models: 1.Core image compression model based on improved SRCNN;2.Core image compression model based on improved ESPCN.In Chapter 4,the basic structure,experimental parameter settings,experimental data,experimental environment and experimental results of the two improved models are introduced in detail.The experimental results are described and analyzed in detail,and the experimental data is used to confirm the analysis.The correctness of the conclusion.The fourth chapter is the performance comparison of the two improved models,including the comparison of the results of the 1x resolution reconstruction,using the compression ratio,SSIM index,PSNR index and visual effects to measure the quality of the reconstructed image.The results of multi-fold(2/3/4)resolution reconstructed images were evaluated using the required storage space size and visual effects as indicators.The results of reconstruction 2/3/4x resolution are compared and analyzed.The experimental results show that the core image compression model based on the improved SRCNN has better results for reconstructing the 1/2/3x resolution image.At this time,the high-definition enlarges the original image and reduces the storage space occupied by the reconstructed image,especially when reconstructing the image with the same size as the original image,the higher compression ratio(2.87-6.06)can be achieved.The improved ESPCN core image compression model has excellent subjective visual effects for reconstructing core images of different magnifications,while the compression ratio is smaller when reconstructing 1x resolution than the improved SRCNN core image compression model.The compression ratio and compression ratio are between 2.98 and 6.93.The only drawback of this model is that as the magnification increases,the amount of storage space required for the reconstructed core image will also increase exponentially.The two core image compression models mentioned in this paper have their own advantages.If only considering the amount of storage space occupied,it is recommended to compress the core image using the improved SRCNN core image compression model;if only the high resolution reconstruction of the image is considered,it is recommended to use the improved ESPCN core image for more compression.The model reconstructs the core image;considering the storage space and high definition of the core image,it is recommended to reconstruct the core image with 2 times resolution using the core image compression model based on the improved ESPCN.Finally,the innovations of this paper are as follows:1.For the first time,the method of image compression based on deep learning and super-sharp image reconstruction is applied to the compression of core images and the reconstruction of core high-definition images.2.After the paper The core image reconstructed by the two improved core image compression models takes up less storage space,has the same visual effect as the original image,and has no defects such as blockiness and edge blur.