Research On Image Coding Algorithm Based On Image Dataset In Wavelet Transform Domain

With the wide popularization of mobile phones,digital cameras and other image capturing terminals,as well as the rapid development of communication technology and image processing technology,the amount of existing image data is growing exponentially.The extensively growing number of images brings great pressure to data storage and transmission.Meanwhile,it may also bring some correlations between the newly added image and the stored images with the increasing number of the stored data.Therefore,if the stored image data set can be used to assist in image coding,the compression efficiency is supposed to be enhanced.In order to make full use of the stored image data sets in the process of image coding and explore the performance benefits of data sets for image coding,this thesis proposes three image coding algorithms based on image data sets in wavelet transform domain,starting from the feature-based image coding algorithm.The main contributions and innovative work of this thesis are summarized as the following three parts:1.A feature-based image coding algorithm in wavelet transform domain is proposed,which uses image features as a bridge and uses similar images stored in the decoder to assist in high-frequency reconstruction.In the encoding process,the input image is decomposed into low-frequency sub-band and high-frequency sub-bands by wavelet transform which are processed separately: the low-frequency sub-band is compressed with high fidelity to retain the global features of the input image;the high-frequency sub-bands are extracted with local features,and the high-frequency sub-bands are transformed from pixel description to feature description,and the obtained feature descriptor is encoded and transmitted.When encoding feature descriptors,the proposed algorithm obtains the feature codebook from the large-scale feature data set training by vector quantization,and maps the high-dimensional descriptors to one-dimensional code-vector index by using the feature codebook,so as to improve the compression efficiency of the descriptors.In the decoding process,the image dataset saved at the decoder side is used to help the reconstruction.Firstly,the decoded low frequency image is used to retrieve the similar images in the image dataset.Then,the feature-codebook is used to reconstruct the high frequency features,which are used for local feature matching to find the similar blocks between the input image and the dataset.Finally,the similar blocks are geometric corrected according to the corresponding positions between matching features,and the corrected similar blocks are used to reconstruct the high frequency sub-bands.After the reconstruction of the high-frequency feature areas one by one,the decoded low-frequency sub-band is combined to perform inverse wavelet transform to get the final reconstructed image.Thus the efficient reconstruction at low bit rate is realized.Experimental results show that the proposed algorithm can achieve a PSNR improvement of 4.55 dB and a SSIM improvement of 0.08 compared with JPEG,it can also achieve a PSNR improvement of 0.18 dB and a SSIM improvement of 0.001 compared with JPEG2000,under the same compression ratio.Furthermore,the proposed algorithm can achieve a better visual reconstruction.2.An image coding algorithm based on dataset prediction in wavelet transform domain is proposed.Starting from the idea of inter-image prediction,the shared image dataset in both encoder and decoder is used for image coding based on prediction.At the encoder side,the low frequency sub-band of the input image is used as the index image to search for the similar image is in the image data set.The retrieved high correlation image is taken as the reference image,and the high frequency sub-bands of the input image is inter-image predicted in blocks.In the phase of image alignment,an image alignment method based on scale invariant feature transform(SIFT)flow pyramid is proposed,and dense features are aligned layer by layer according to the feature matching relationship to achieve more accurate image alignment.In the phase of image prediction in blocks,a similarity threshold is set to adaptively judge the prediction accuracy of each block,and the blocks which cannot be correctly predicted are encoded as prediction residual to reduce the reconstructive distortion caused by the prediction error.At the decoder side,the decoded low frequency image is firstly used for HCOI retrieval in the image dataset.Then the corresponding high frequency blocks are extracted from the aligned HCOIs according to the prediction information.High frequency predictions,together with the decoded prediction residuals,are used to form the reconstructed high frequency sub-bands.Finally,the decoded low frequency sub-band and the reconstructed high frequency sub-bands are inverse transformed to obtain the reconstructed image.The proposed method utilizes the shared image dataset to do the inter-frame prediction,which can remove the inter-frame redundancy and increase the coding efficiency of the high frequency sub-bands.Furthermore,thanks to the adaptive residual selection strategy,the proposed method can perform well reconstruction even if there is no HCOI existed in the shared image dataset.Experimental results show that,when encoding the test images,the proposed algorithm shows excellent rate-distortion performace especially at low bit rates.Under the same compression ratio,the proposed algorithm can achieve a maximum PSNR improvemet of 7dB compared with JPEG,and it can achieve a maximum PSNR improvement of 1.69 dB compared with JPEG2000,furthermore,it can achieve a maximum MS-SSIM improvement of 0.03 compared with HEVC intra coding.3.A deep image coding algorithm based on high frequency sub-band prediction in wavelet transform domain is proposed.A large-scale image database is used as training set to train the image coding model based on deep convolutional auto-encoder.After the wavelet transform,four wavelet sub-bands are encoded and decoded separately by four parallel neural networks,and both the encoder and the decoder in each branch network are fully convolutional.At the encoder side,multi-scale features of sub-bands are extracted by the convolutional layers with different strides.Correspondingly,at the decoder side,transposed convolutional layers with different strides are used to restore the sub-band data from the characteristics of the codeword space.In order to remove the redundancy between sub-bands before encoding,a high-frequency prediction network is proposed and designed to obtain the prediction of high-frequency sub-bands from low-frequency sub-band,so that only the high frequency residuals need to be encoded.In addition,in order to further improve the coding efficiency,the conditional probability-based entropy coding model is used to estimate the prior probability of the codes.Thereby,the bit rate estimation can be further obtained to realize the jointly rate-distortion optimization in the training process.Experimental results show that the proposed coding networks can achieve a better rate-distortion performance compared with both traditional coding methods and classic deep image coding methods,and it also shows obvious advantages in the visual comparion.Compared with the recently proposed deep image coding methods,the propsed method can achieve 0.0048 MS-SSIM improvement in high frequency reconstruction.Furthermore,the proposed prediction model can bring an average MS-SSIM improvement of 0.0019 compared with the coding model without prediction,it can also effectively eliminate the jagged edges in the reconstructed image,which can make a better high frequency reconstruction.