Image compression reconstruction method based on vector quantization

Abstract

The application relates to the technical field of image compression, and discloses an image compression and reconstruction method based on vector quantization, which compresses and reconstructs an input image x through an image coding and decoding model to obtain a reconstructed image, the image coding and decoding model comprises a multilayer structure; each layer structure comprises a down-sampling layer, an up-sampling layer and a plurality of quantization layers; the down-sampling layers and the quantization layers in the layer structures are sequentially arranged to form a compression path; the up-sampling layers and the quantization layers in the layer structures are sequentially arranged to form a reconstruction path; compared with a traditional coding architecture, the upper bound of rate distortion performance is higher, the problem of excessively high coding and decoding complexity of vector quantization is solved, and the practicability of the whole framework is improved.

Claims

1. An image compression and reconstruction method based on vector quantization, characterized in that, The input image is processed through an image encoding and decoding model. Compression and reconstruction are performed to obtain the reconstructed image. : The image encoding and decoding model includes a multi-layer structure; each layer includes a downsampling layer, an upsampling layer, and several quantization layers; the downsampling layers and quantization layers in each layer are sequentially combined to form a compression path; The upsampling layer and quantization layer in each layer structure are sequentially combined to form the reconstruction path; The image x is downsampled to obtain the vector to be transformed, which is then input to the Lth quantization layer. Each quantization layer transforms the input vector to be transformed. The transformed vector obtained by performing vector transformation , and the The reconstructed vector output by each quantization layer Subtraction and linear mapping yield the vector to be quantized. , , This represents the total number of quantization layers in the image encoding / decoding model. The shape is , The size of the channel dimension. For spatial resolution, For spatial height, For the space width, each The channel tensor is treated as a subvector; the quantizer is constrained by entropy through multiple subvectors. In the vector to be quantized Multi-level vector quantization is performed to obtain the quantized vector. , After a linear mapping and Adding them together yields a vector. , will vector Perform vector transformation to obtain the first The reconstructed vector output by each quantization layer The reconstructed vector output from the Lth quantization layer As a reconstructed image Specifically, it includes: In each quantization stage, the vector to be quantized is decomposed into multiple sub-vectors, and each sub-vector is quantized separately. Vector transformation is performed in the quantization layer using VT units. Each VT unit consists of cascaded intra-vector and inter-vector transformations. The intra-vector transformation comprises two intra-channel transformation layers and a nonlinear activation layer located between them. The intra-channel transformation layers are constructed using neural networks. The inter-vector transformation includes a channel-by-channel intra-block transformation layer, where the channel-by-channel intra-block transformation layer transforms the vector to be transformed at each dimension of... Inside the block The subvectors are transformed channel by channel, with different channels having different transformation matrices.

2. The image compression and reconstruction method based on vector quantization according to claim 1, characterized in that: Quantizer constrained by subvector entropy vector to be quantized When performing quantitative reconstruction, Send in Subvector entropy-constrained quantizer Multi-level quantization reconstruction is performed to obtain the quantized vector. : vector ,in for sub-vector, the first The reconstructed vector output by each quantization layer , for The subvector; then the first The vector to be quantized in each quantization layer is ;right middle Subvectors are used Subvector entropy-constrained quantizer Quantization is performed separately to obtain the quantized vector. ; It is a linear mapping.

3. The image compression and reconstruction method based on vector quantization according to claim 1, characterized in that: The loss function F of the image encoding and decoding model is: ； in, For codeword index, For mathematical expectation symbol, As a measure of spatial distance to information sources, This is a probabilistic model for codeword indexing. To transform the spatial distance metric, For distortion balance control rate, control and The balance between them.

Images