Half-tone image super-resolution reconstruction method and system

Abstract

This invention discloses a method and system for super-resolution reconstruction of halftone images. The method includes: multi-scale feature extraction and fusion of a low-resolution continuous-tone image; enhancement of spatial features through variable large-kernel attention; decomposition of frequency domain features using dual-tree complex wavelet transform; and feature fusion and resolution enhancement through local linear multi-head self-attention to obtain a high-resolution continuous feature map. Subsequently, differentiable discretization processing and visual perception optimization are performed using the Gumbel-Softmax mechanism to obtain a high-resolution halftone image. This invention, through the combined effects of spatial and frequency domain co-modeling, near-linear complexity attention computation, and end-to-end differentiable discretization decision-making, significantly reduces computational overhead and memory usage while effectively improving the detail realism, orientation consistency, and visual quality of the reconstructed image. It is suitable for real-time printing scenarios with low latency and low power consumption at the edge.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for super-resolution reconstruction of halftone images. Background Technology

[0002] Digital halftone and printed image reconstruction technology has evolved through several important stages since its early stages, including thresholding, random dithering, and error diffusion (ED). These stages include blue noise templates, model-driven direct binary search (DBS), multitone / variable drop halftone, and quality optimization incorporating visual models (such as the human visual system's HVS, dot spread function, and dot gain). With the continuous evolution of industrial inkjet printing technology in terms of resolution, paper speed, printhead density, and variable droplet capability, the front-end raster image processor (RIP) link has gradually expanded from simple color management and dot generation to a holistic optimization scheme that integrates device physical constraints and spectral characteristics. However, the traditional "two-stage" processing flow (i.e., interpolation or super-resolution reconstruction followed by halftone processing) has gradually revealed its inherent limitations in terms of high-frequency detail preservation, moiré suppression, link bandwidth, and end-side computation latency.

[0003] In recent years, the development of deep learning technology has greatly promoted the integration of image super-resolution (SR) and halftone reconstruction tasks. From early SRCNN and VDSR, to methods based on generative adversarial networks (GANs) and perceptual loss, and more recently lightweight attention mechanisms and large-kernel convolutional designs, the focus of the technology has shifted from purely "spatial domain" processing to "frequency domain and directionality co-modeling." At the same time, industry applications have placed stringent requirements on the real-time performance, low power consumption, and small memory footprint of edge devices, making deep separable convolutions, reparameterization techniques, INT8 quantization, and windowing / linear attention mechanisms the mainstream lightweight engineering methods.

[0004] Despite continuous technological advancements, current industrial production processes still predominantly employ traditional interpolation algorithms (such as bitrilinear interpolation and bicubic interpolation) combined with error diffusion or DBS halftone methods. While some attempts have attempted to introduce learning-based super-resolution models to improve edge and texture reconstruction quality, limitations in the non-differentiable quantization process, high transmission bandwidth requirements, and limited computational power at the edge side mean that achieving end-to-end high-resolution halftone image generation remains a technological transitional phase. Furthermore, the insufficient coverage of existing public datasets and benchmarks in real-world printing conditions (such as PASS strategies, dot gain effects, and differences in printing media) further restricts the horizontal comparison and reproducibility of different methods in industrial scenarios.

[0005] Existing related technologies have many drawbacks. For example, while image super-resolution reconstruction methods based on bilinear interpolation can enlarge image size, they cannot effectively recover high-frequency details, resulting in blurred and distorted reconstructed images. Classical error diffusion halftone methods are prone to producing stripe artifacts and noise when processing complex images, and their high computational complexity makes them unsuitable for real-time processing. Direct Binary Search (DBS) methods can generate high-quality binary images, but their huge computational cost and search optimization process result in extremely low algorithm efficiency, making them unsuitable for high-resolution images or scenarios with high real-time requirements. Recently, some research has attempted to achieve super-resolution of color halftone images using convolutional neural networks (such as EDSR) and proposed Halftone-Net combined with Fast Fourier Transform (FFT) to suppress moiré patterns. However, these methods still suffer from high computational cost, insufficient image quality, and incomplete moiré pattern elimination, limiting their feasibility and widespread adoption in industrial applications. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problems of severe loss of image details, large bandwidth requirements for transmission of intermediate high-resolution images, and low computational energy efficiency in the prior art.

[0007] To address the aforementioned technical problems, this invention provides a halftone image super-resolution reconstruction method, comprising the following steps: S1: Perform multi-scale feature extraction and fusion on low-resolution continuous-tone images, and improve spatial resolution to obtain high-resolution continuous feature maps; S2: Discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution. S3: Perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

[0008] In one embodiment of the present invention, step S1, the method for performing multi-scale feature extraction and fusion on the low-resolution continuous-tone image and improving its resolution, includes the following steps: S1.1: Perform initial feature mapping on the low-resolution continuous-tone image to extract a primary feature map containing basic features; S1.2: Spatial feature enhancement is performed on the primary feature map using variable large kernel attention to obtain an intermediate feature map with enhanced global context information; S1.3: Perform frequency domain feature decomposition on the intermediate feature map to obtain multi-subband frequency domain features; S1.4: The multi-subband frequency domain features are fused by local linear multi-head self-attention to obtain deep enhanced features that have been optimized by spatial and frequency domains. S1.5: Improve the spatial resolution of the deep enhancement features to obtain a high-resolution continuous feature map.

[0009] In one embodiment of the present invention: spatial feature enhancement is performed on the primary feature map using a variable large kernel attention function, wherein the variable large kernel attention function is: , in, Let be a variable large kernel attention function, representing the intermediate feature map output by the variable large kernel attention function that enhances global contextual information. This represents a 1x1 point convolution. Represent a Depth-separable convolutions with variable kernel size. Represent a Depth-dilated convolution with variable kernel size. For variable kernel attention, the kernel size decreases along the network depth, and X is the input primary feature map.

[0010] In one embodiment of the present invention, in step S1.3, the method for performing frequency domain feature decomposition on the intermediate feature map is as follows: performing a dual-tree complex wavelet forward transform on the input intermediate feature map to decompose the spatial domain features into the frequency domain, obtaining one low-frequency approximate sub-band and six high-frequency detail sub-bands in different directions; the six high-frequency detail sub-bands in different directions correspond to ±15°, ±45°, and ±75° directions, respectively; the low-frequency approximate sub-band and the six high-frequency detail sub-bands constitute the multi-sub-band frequency domain features.

[0011] In one embodiment of the present invention, in step S1.4, the method of feature fusion of the multi-subband frequency domain features by local linear multi-head self-attention is as follows: local linear multi-head self-attention divides the input multi-subband frequency domain features into multiple non-overlapping fixed-size windows in the spatial dimension; within each window, the multi-subband frequency domain features are aggregated and enhanced by a linear multi-head attention mechanism, and the output features of all windows are fused to generate deep enhanced features after spatial and frequency domain co-optimization.

[0012] In one embodiment of the present invention, step S2 involves discretizing the high-resolution continuous feature map to obtain a discrete ink level probability distribution. The method for generating a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution includes the following steps: S2.1: Perform a linear transformation on the continuous feature vector at each pixel position in the high-resolution continuous feature map to output a K-dimensional feature vector, where K is the number of discrete ink levels; S2.2: The input probability of the K-dimensional feature vector is converted into a discrete ink level probability distribution by the Softmax function, which represents the probability of each pixel selecting a different ink level; S2.3: Generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution.

[0013] In one embodiment of the present invention, in step S2.3, the method for generating a preliminary halftone image of pixel-level discrete ink dot distribution based on the discrete ink level probability distribution is as follows: during the training phase, Gumbel-Softmax sampling is used to perform differentiable discrete decision on the discrete ink level probability distribution; during the inference phase, argmax operation is used to perform hard decision on the discrete ink level probability distribution to generate a preliminary halftone image of pixel-level discrete ink dot distribution.

[0014] In one embodiment of the present invention, in step S3, the method for visual perception optimization of the preliminary halftone image is as follows: the preliminary halftone image is filtered in the frequency domain to suppress moiré noise at a specific frequency; an anisotropic diffusion algorithm is used to smooth artifacts in the image edge region and output a high-resolution halftone image.

[0015] Based on the same inventive concept, the present invention also provides a halftone image super-resolution reconstruction system, which includes the following modules in sequence: a super-resolution reconstruction module, used to extract and fuse multi-scale features of low-resolution continuous tone images and improve spatial resolution to obtain a high-resolution continuous feature map; The halftone generation module is used to discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution. The visual perception optimization module is used to perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

[0016] The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the halftone image super-resolution reconstruction method described above.

[0017] The technical solution of the present invention has the following advantages compared with the prior art: The method described in this invention integrates super-resolution reconstruction and halftone generation modules into a single, end-to-end lightweight network design, effectively avoiding the detail loss and transmission overhead issues inherent in traditional two-stage processing. Through innovative attention mechanisms and frequency domain transformation techniques, it significantly improves image detail recovery and visual quality. The lightweight design drastically reduces computational complexity, enabling the model to meet the real-time processing requirements of edge devices. Simultaneously, differentiable discrete decision training and efficient inference mechanisms significantly improve processing efficiency while maintaining image quality. The final visual optimization module further suppresses artifacts, enhances the visual consistency and printability of halftone images, and comprehensively improves the practical value of the method. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating a halftone image super-resolution reconstruction method provided in an embodiment of the present invention. Figure 2 This is a comparative schematic diagram of large kernel attention and variable large kernel attention in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the operation of the dual-tree wavelet transform module in an embodiment of the present invention. Figure 4 This is a schematic diagram of a local linear multi-head self-attention framework in an embodiment of the present invention; Figure 5 This is a schematic diagram of the Gumbel-Softmax processing flow in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the super-resolution reconstruction effects of different algorithms at a preset scale in Experiment 1 of this invention; Figure 7 This is a first comparative diagram of the effects of different halftone algorithms in Experiment 1 of this invention; Figure 8 This is a second comparative diagram showing the effects of different halftone algorithms in Experiment 1 of this invention; Figure 9 This is a schematic diagram of the actual operation process in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0021] Example 1: like Figure 1As shown, the present invention provides a method for super-resolution reconstruction of halftone images, comprising the following steps: S1: Perform multi-scale feature extraction and fusion on low-resolution continuous-tone images, and improve spatial resolution to obtain high-resolution continuous feature maps; S2: Discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution. S3: Perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

[0022] The halftone image super-resolution reconstruction method provided by this invention first performs multi-scale feature extraction and fusion on the input low-resolution continuous-tone image, and then improves its spatial resolution through upsampling to obtain a high-resolution continuous feature map containing rich details. The purpose of this step is to effectively recover high-frequency information and global context from low-quality input. Subsequently, the obtained high-resolution continuous feature map is discretized to generate a discrete ink-level probability distribution, and a preliminary halftone image with pixel-level discrete ink dot distribution is generated based on this distribution, thereby converting the continuous-tone features into a discrete ink dot representation suitable for printing. Finally, the preliminary halftone image is visually optimized to further improve subjective quality by suppressing moiré patterns and edge artifacts, ultimately outputting a clear and natural high-resolution halftone image.

[0023] This invention employs an end-to-end deep learning framework to achieve super-resolution reconstruction of halftone images. The method mainly comprises two core components: a super-resolution reconstruction module based on lightweight attention (SR-NET) and a halftone generation module based on Gumbel-Softmax ink dot control (HT-NET).

[0024] In the SR-NET module, multi-scale features are first extracted from the input low-resolution continuous grayscale image through shallow and deep feature extraction modules. These features are then fused through a multi-layer feature aggregation module to fully capture the complex features and high-frequency information of the image. After feature fusion, a pixel-shuffle upsampling operation is used to map the fused features to a high-resolution image. To further reduce computational complexity and improve deployment efficiency, a depthwise separable convolutional structure with variable kernel size is introduced, achieving a lightweight modification to the network.

[0025] In the HT-NE module, local ink dot distribution control is performed on the high-resolution continuous feature map output by SR-NET. Gumbel-Softmax sampling is employed to achieve differentiable discretization. During the training phase, gradient optimization ensures the preservation of image details and quality, while in the inference phase, deterministic decisions are made, ultimately generating a high-quality halftone image. This method achieves high-quality conversion from low-resolution grayscale input to high-resolution halftone output through end-to-end optimization.

[0026] To achieve the goal of strong detail recovery and high computational efficiency in image super-resolution reconstruction, this invention proposes an SR-NET module based on lightweight self-attention and frequency domain enhancement. This module integrates variable big kernel attention mechanism (VLKA), dual-tree complex wavelet transform (DTCWT) assisted feature enhancement, local linear multi-head self-attention mechanism (LSA), and parameterized depthwise separable convolutional structure, which significantly reduces computation and storage overhead while improving reconstruction quality.

[0027] A variable large kernel attention module is used to establish long-range dependencies with near-linear computational complexity. This module performs attention weight calculation within a large receptive field by grouping convolutions and generating dynamic kernels on the input primary feature maps, thus avoiding the secondary complexity problem present in traditional self-attention mechanisms. Building upon this, a dual-number complex wavelet transform is introduced to enhance the features in the frequency domain, decomposing the image into low-frequency and high-frequency sub-bands. Low-frequency information is used to maintain structural consistency, while high-frequency information is enhanced through dedicated convolutions, thereby improving the reconstruction quality of edge and texture details.

[0028] To further reduce memory usage and latency during attention computation, this invention employs a local linear multi-head self-attention framework. This framework reduces the computational scale of query-key-value pairs through linear projection and a local window mechanism, minimizing memory access overhead while maintaining non-local modeling capabilities. Simultaneously, to achieve lightweight and deployment-friendly modeling, the convolutional module utilizes a reparameterized, depthwise separable convolutional structure. During training, multi-branch convolutions enhance feature representation, while during inference, the structure is reparameterized into a single convolutional layer, balancing performance and inference efficiency while improving quantization adaptability.

[0029] Specifically, in step S1, the method for multi-scale feature extraction and fusion of the low-resolution continuous-tone image and for improving its resolution includes the following steps: S1.1: Perform initial feature mapping on the low-resolution continuous-tone image to extract a primary feature map containing basic features; S1.2: Spatial feature enhancement is performed on the primary feature map using variable large kernel attention to obtain an intermediate feature map with enhanced global context information; S1.3: Perform frequency domain feature decomposition on the intermediate feature map to obtain multi-subband frequency domain features; S1.4: The multi-subband frequency domain features are fused by local linear multi-head self-attention to obtain deep enhanced features that have been optimized by spatial and frequency domains. S1.5: Improve the spatial resolution of the deep enhancement features to obtain a high-resolution continuous feature map.

[0030] In step S1.1, for the input low-resolution continuous tone input image, an initial feature mapping is performed through a convolution operation to extract a primary feature map containing basic structure and texture information, providing basic feature representation for subsequent deep feature extraction and reconstruction.

[0031] Furthermore, such as Figure 2 As shown, in step S1.2, after obtaining the initial feature map, feature enhancement is performed using VLKA. Unlike traditional fixed large kernel attention (LKA), which uses a single kernel size K in each layer, VLKA uses a different kernel size. To obtain a larger receptive field, K needs to be increased or the number of layers increased, which can easily lead to computational or parameter redundancy and a limited response to high-frequency details. VLKA uses a decreasing kernel size along the network depth. It rapidly expands the global receptive field in shallow layers, suppresses redundancy in deep layers, and significantly reduces the nuclear area at a similar equivalent receptive field. This results in increased floating-point operations (FLOPs), parameter count, and memory bandwidth overhead.

[0032] VLKA decomposes traditional large-kernel convolution into three consecutive operations for efficient computation. It uses 1×1 point convolutions to map and integrate input features along the channel dimension; and employs depthwise separable convolutions with variable kernel sizes to extract spatial features. Where d is a hyperparameter controlling the kernel size; the receptive field is further expanded using depthwise dilated convolution with a variable kernel size, and the kernel size is... This process can be represented as: , in, Let be a variable large kernel attention function, representing the intermediate feature map output by the variable large kernel attention function that enhances global contextual information. This represents a 1x1 point convolution. Represent a Depth-separable convolutions with variable kernel size. Represent a Depth-dilated convolution with variable kernel size. For variable-kernel attention, the kernel size decreases along the network depth, where X is the input primary feature map. That is, it is a real matrix with C channels, a height of H, and a width of W.

[0033] Furthermore, such as Figure 3 As shown, in step S1.3, the intermediate feature map obtained in step S1.2 is processed. Perform dual-tree complex wavelet forward transform The spatial domain features are mapped to the frequency domain and decomposed into a low-frequency approximate sub-band. and 6 high-frequency detail subbands in different directions These six high-frequency subbands correspond to ±15°, ±45°, and ±75° directions, respectively, enabling the capture of multi-directional edge and texture structures. The low-frequency subbands carry image smoothing and contour information, while the high-frequency subbands contain detail components. Together, the low-frequency and high-frequency subbands constitute a multi-scale, multi-directional frequency domain feature representation. The mathematical formula for this decomposition process is: , Among them, low-frequency components and high frequency components Each is composed of complex wavelet coefficients and Through wavelet basis functions and direction sensitive It is composed of linear combinations, and the specific formula is as follows: , , in, Indicates scale. Represents the set of directions .

[0034] The real and imaginary coefficients used in the transformation are generated by two sets of discrete filter banks, which together provide approximate four-phase orthogonality and translation invariance.

[0035] The dual-number complex wavelet transform is implemented using a convolutional filter bank. Its computational complexity is nearly linearly related to image resolution, and it has low memory consumption, making it suitable for deployment on edge computing devices. The inverse dual-number complex wavelet transform (IDTCWT) can reconstruct spatial features losslessly or nearly losslessly based on the aforementioned subbands. The reconstruction process is expressed by the following formula: .

[0036] The frequency domain decomposition module complements the variable big kernel attention mechanism. The frequency domain decomposition module focuses on long-range dependency modeling in the spatial domain, while the variable big kernel attention explicitly enhances high-frequency details and structural consistency, significantly improving the clarity and stability of the reconstructed image with almost no increase in computational overhead.

[0037] Furthermore, such as Figure 4As shown, in step S1.4, the multi-subband frequency domain features are fused through a local linear multi-head self-attention mechanism to achieve the purpose of synergistic optimization of spatial and frequency domain information under approximately linear computational complexity, thereby obtaining deep enhanced features.

[0038] Local linear multi-head self-attention divides the input multi-subband frequency domain features into multiple non-overlapping fixed-size windows in the spatial dimension; within each window, the multi-subband frequency domain features are aggregated and enhanced through a linear multi-head attention mechanism, and the output features of all windows are fused to generate deep enhanced features after spatial and frequency domain co-optimization.

[0039] Specifically, the local linear multi-head self-attention mechanism uses multi-subband frequency domain features As input, multi-subband frequency domain features The value after VLKA encoding and normalization is represented as V, while the frequency domain feature A extracted by DTCWT contains multi-directional and phase information; both are dimensional. The tensor.

[0040] Rearrange A and V by the number of heads. ,in The total number of spatial pixels, calculated according to the preset window size. Divided into By dividing the non-overlapping blocks, a sequence is obtained. .

[0041] Linear attention computation is performed within each block, approximating the traditional attention mechanism through element-wise multiplication and normalization operations, thus avoiding the explicit construction of a quadratic complexity attention matrix. The process is expressed by the following formula: , in, This indicates element-wise multiplication. This indicates a normalization operation.

[0042] Output of each block The feature maps are stitched together along the spatial dimensions to form a complete feature map. Add convolution position Then, through residual connections and 1×1 convolution to integrate the features, the deep enhancement features that fuse the KYC and frequency domain features are finally obtained. The formula for this process is expressed as: , , in, Represents the normalization function. This represents the intermediate output features of a local linear multi-head self-attention mechanism. This indicates that 1×1 convolution is used to integrate and transform the channel dimensions. This represents the final output of local linear multi-head self-attention, i.e., deep enhancement features.

[0043] This locally linear multi-head self-attention mechanism reduces computational and memory complexity from that of traditional self-attention mechanisms. Reduced to approximately linear It significantly reduces memory consumption and bandwidth pressure. At the same time, by working in conjunction with the long-range dependency modeling capabilities provided by VLKA and the high-frequency detail enhancement mechanism of DTCWT, it improves the consistency of reconstruction details while meeting the deployment requirements of edge devices for low latency, low power consumption and small memory consumption.

[0044] Furthermore, in step S1.5, the spatial resolution of the deep enhancement feature is improved.

[0045] Optionally, this can be achieved through a pixel shuffle operation, which reorganizes the channel dimension information into the spatial dimension to obtain a high-resolution continuous feature map.

[0046] This upsampling method integrates feature information in a learnable manner, enhancing the recovery quality of high-frequency details while magnifying the image, and providing high-resolution, continuous-value feature representations for subsequent quantization or reconstruction operations.

[0047] Furthermore, in the HT-NE module, local ink dot distribution control is performed on the high-resolution continuous feature map output by SR-NET. In step S2, the high-resolution continuous feature map is discretized to obtain a discrete ink level probability distribution. The method for generating a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution includes the following steps: S2.1: Perform a linear transformation on the continuous feature vector at each pixel position in the high-resolution continuous feature map to output a K-dimensional feature vector, where K is the number of discrete ink levels; S2.2: The input probability of the K-dimensional feature vector is converted into a discrete ink level probability distribution by the Softmax function, which represents the probability of each pixel selecting a different ink level; S2.3: Generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution.

[0048] Specifically, in step S2.1, the high-resolution continuous feature map is processed. The continuous feature vectors at each pixel location are linearly transformed, and then a lightweight classification head consisting of one or more 1×1 convolutional layers is used to map the C-dimensional features of each pixel into a K-dimensional feature vector, where K is the preset number of discrete ink levels, and the vector elements represent the unnormalized log-probability of the corresponding ink level. .

[0049] Furthermore, in steps S2.2 and S2.3, the K-dimensional feature vector is converted into a discrete ink level probability distribution using the Softmax function, representing the normalized probability of each pixel selecting different ink levels. Based on the discrete ink level probability distribution, a preliminary halftone image with pixel-level discrete ink dot distribution is generated, specifically including two stages: training and inference.

[0050] like Figure 5 As shown, during the training phase, Gumbel-Softmax sampling is used to achieve a differentiable discrete decision approximation, and independent and identically distributed Gumbel noise is added to each ink level of each pixel. Then according to temperature parameters Normalization is performed using the following formula: , The training uses a Straight-Through estimator, and the forward propagation uses hard one-hot decision, as shown in the formula: , Backpropagation uses the gradient of a soft probability distribution and the temperature parameter. As training progresses, the sample is gradually annealed, decreasing from a high value to near zero, so that the sampling results smoothly transition from soft probabilities to hard decisions.

[0051] During the inference phase, noise addition and the Softmax operation are omitted. Instead, the ink level with the highest logarithmic probability is directly selected for each pixel to generate a deterministic discrete output. , This yields a preliminary halftone image with pixel-level discrete ink dot distribution.

[0052] Further, in step S3, the preliminary halftone image is visually optimized to obtain a high-resolution halftone image. The preliminary halftone image is then filtered in the frequency domain to suppress moiré noise at a specific frequency. An anisotropic diffusion algorithm is used to smooth artifacts in the image edge region, and a high-resolution halftone image is output.

[0053] This module uses Gumbel-Softmax to make pixel-level discrete quantization into forward hard and backward soft differentiable sampling, allowing the selection of halftone points or levels to be learned directly in an end-to-end framework, avoiding the quality bottlenecks and time consumption caused by traditional two-stage methods. Temperature annealing ensures stable training and gradually approximates hard decision-making. Combined with anisotropy and frequency domain constraints, it can better preserve edges and textures and reduce moiré patterns. In the inference stage, noise is removed and argmax is taken, resulting in deterministic output, low latency, and small memory usage. The operators are all convolution and element-wise operations, making quantization and deployment friendly. At the same time, it supports expansion from two levels to multiple levels and from single channel to multiple channels, significantly improving efficiency and deployment usability while maintaining visual quality close to DBS.

[0054] Furthermore, the training process of this invention employs an end-to-end optimization method based on a comprehensive loss function, aiming to collaboratively optimize the continuous domain super-resolution reconstruction and discrete halftone generation modules. The loss function is designed to closely align with the "print-observe" perception link and imposes constraints in the spectral and spatial domains to improve visual quality. Training comprises two phases with different focuses, sharing network weights and being optimized through alternating or joint iterations.

[0055] Phase 1 is continuous grayscale supervision, using continuous grayscale images. For the objective, the halftone result output by the network is denoted as . Diffusion of nuclei using Gaussian points Approximate HVS involves restoring the discrete point lattice to the receptive domain, calculating the consistency between the reconstruction and the structure, and introducing a consistency term with the DBS label: , , , in For the corresponding DBS halftone labels. The total loss for stage one is: .

[0056] Phase two involves anisotropic learning, using a fixed grayscale value of 0-255 for the image. For input, network output Further minimizing the error in the receptive domain, and incorporating anisotropic constraints to learn the directional point distribution: , , The total loss in Phase Two is: .

[0057] The two-stage shared network weights, alternating or jointly iterating, have the overall optimization objective as follows: .

[0058] Experiment 1: To fully demonstrate the amplification effectiveness of this invention, the superiority of the SR-NET module in super-resolution tasks was tested, and the effects of end-to-end networks and traditional halftone processing methods were compared.

[0059] like Figure 6 As shown, to verify the effectiveness of the SR-NET module in amplification reconstruction and engineering implementation, a comparative experiment was conducted on the lightweight self-attention super-resolution reconstruction module. The more challenging ×4 amplification was selected as the evaluation scenario, and it was compared with several super-resolution networks that currently perform better.

[0060] Experimental results show that the SR-NET module performs exceptionally well in high-frequency texture and detail restoration, with clearer edges and more continuous textures. Overall, both subjective and objective indicators are superior to the comparison methods, providing a more refined detail foundation and a more stable reconstruction prior for subsequent super-resolution reconstruction of halftone images.

[0061] To verify the effectiveness of the proposed end-to-end halftone network, a systematic comparison was conducted between this method and traditional halftone workflows (interpolation / super-resolution followed by error diffusion or DBS) in terms of image reconstruction quality and processing efficiency. The evaluation was performed under identical data and hardware conditions: the inputs were all low-resolution images of the same resolution; the traditional workflow first generates a continuous high-resolution image in the middle, and then performs ED or DBS at that resolution; this invention directly outputs high-resolution halftone images at the edges in a single pass from the low-resolution input. To closely reflect practical viewing and printing, this comparison evaluated HVS-PSNR or HVS-SSIM, as well as edge-side inference latency, CPU utilization, and peak memory usage in the "print-observation" simulation domain.

[0062] like Figure 7 and Figure 8 As shown in the qualitative results, the halftone details and textures generated by this invention are highly consistent with the DBS results, with sharp edge contours and natural transitions between fine grains. Compared to error diffusion, it significantly reduces the striping and blocky aggregation effects caused by low-to-mid-frequency clustering, and is closer to the "blue noise" distribution in regular textures and high-frequency regions, resulting in a smoother subjective appearance that conforms to human visual characteristics. Frequency domain visualization shows that the energy of this invention is controlled in the sensitive frequency band, and the moiré pattern suppression effect is better; the directional energy distribution is consistent with the expected anisotropic prior, and the directional stability is stronger.

[0063] As shown in Table 1, taking a 500×342 pixel image as an example, quantitative results show that, under the same frequency limiting and thread configuration, this invention has significant advantages in edge-side inference latency and CPU usage, and can generate high-quality halftone output in a shorter time. By avoiding the transmission of intermediate high-resolution images on the link, the end-to-end path of this invention further reduces bandwidth and overall latency. In terms of perceptual domain metrics, HVS-PSNR or HVS-SSIM is superior to the traditional DBS processing flow, with lower moiré index and higher spectral flatness. Anisotropy metrics indicate that this method is more stable in preserving directional texture and detail. Considering both quality and efficiency, this invention achieves significant speedup and reduced resource consumption while maintaining image quality close to that of DBS, better meeting the engineering constraints of real-time edge printing.

[0064] Table 1:

[0065] Example 2: like Figure 9 As shown, in the actual printing scenario, the training process employs a dual-pipeline parallel end-to-end workflow. Pipeline 1 takes a low-resolution continuous grayscale image X and its Gaussian auxiliary channel (generated from X by a Gaussian kernel) as input, feeds it into the feature extraction backbone (i.e., VLKA to DTCWT to LSA), and after pixel-level upsampling (PixelShuffle or upsampling head), it is connected to a Gumbel-Softmax quantization head to generate an enlarged discrete halftone image; this pipeline uses... Constraint Reconstruction and Perceived Quality. Pipeline 2 takes a set of fixed grayscale images (0-255 representative density / grayscale bars) and their Gaussian auxiliary channels as input, uses the same main trunk and quantization head to generate corresponding halftone outputs, and uses... Training emphasizes the learning of anisotropy (directional texture, point spread differences); where the anisotropy term measures texture orientation consistency based on directional gradient or frequency domain (e.g., DTCWT directional energy).

[0066] The two pipelines share weights and perform joint optimization. Gumbel-Softmax employs temperature annealing: during training, forward pass approaches hard selection and backward pass is differentiable, steadily transitioning to discrete decision-making. During the inference phase, the host computer only transmits low-resolution grayscale images, while the edge-end loads trained weights. Inference is performed in a row / block streaming manner, and the high-resolution half-scale image is directly output via the upsampling head and Gumbel-Softmax's argmax. The output can be seamlessly written to the print buffer and integrated with the actual production process (multi-pass masking, nozzle scheduling, dot gain compensation), achieving low-bandwidth transmission, low-latency generation, and predictable ink drop in online printing.

[0067] Example 3:

[0068] Based on the same inventive concept as Embodiment 1, the present invention also provides a halftone image super-resolution reconstruction system for implementing the halftone image super-resolution reconstruction method described in Embodiment 1, comprising the following modules: The super-resolution reconstruction module is used to extract and fuse multi-scale features from low-resolution continuous-tone images and improve spatial resolution to obtain high-resolution continuous feature maps. The halftone generation module is used to discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution. The visual perception optimization module is used to perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

[0069] Specifically, the super-resolution reconstruction module sequentially includes a shallow convolutional layer, a variable large kernel attention module, a dual-tree complex wavelet transform module, a local linear multi-head self-attention module, and a pixel rearrangement upsampling module; The shallow convolutional layer is used to perform initial feature mapping on the low-resolution continuous-tone image and extract a primary feature map containing basic features. The variable kernel attention module is used to perform spatial feature enhancement on the primary feature map through variable kernel attention, so as to obtain an intermediate feature map with enhanced global context information. The dual-tree complex wavelet transform module is used to perform frequency domain feature decomposition on the intermediate feature map to obtain multi-subband frequency domain features. The local linear multi-head self-attention module is used to perform feature fusion on the multi-subband frequency domain features through local linear multi-head self-attention, so as to obtain deep enhanced features after spatial and frequency domain co-optimization. The pixel rearrangement upsampling module is used to improve the spatial resolution of the deep enhancement features to obtain a high-resolution continuous feature map.

[0070] Furthermore, the halftone generation module sequentially includes a feature mapping layer, a probability distribution generation layer, and a discrete decision layer; The feature mapping layer is used to perform a linear transformation on the continuous feature vector at each pixel position in the high-resolution continuous feature map, and output a K-dimensional feature vector, where K is the number of discrete ink levels. The probability distribution generation layer is used to convert the input probability of the K-dimensional feature vector into a discrete ink level probability distribution through the Softmax function, representing the probability of each pixel selecting a different ink level; The discrete decision layer is used to generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution.

[0071] Example 4: The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the halftone image super-resolution reconstruction method described in Embodiment 1.

[0072] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for super-resolution reconstruction of a semi-processed image, characterized in that, Includes the following steps: S1: Perform multi-scale feature extraction and fusion on low-resolution continuous-tone images, and improve spatial resolution to obtain high-resolution continuous feature maps, specifically: S1.1: Perform initial feature mapping on the low-resolution continuous-tone image to extract a primary feature map containing basic features; S1.2: Spatial feature enhancement is performed on the primary feature map using variable large kernel attention to obtain an intermediate feature map with enhanced global context information; S1.3: Perform frequency domain feature decomposition on the intermediate feature map to obtain multi-subband frequency domain features; S1.4: The multi-subband frequency domain features are fused by local linear multi-head self-attention to obtain deep enhanced features that have been optimized by spatial and frequency domains. S1.5: Improve the spatial resolution of the deep enhancement features to obtain a high-resolution continuous feature map; S2: Discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution, specifically: S2.1: Perform a linear transformation on the continuous feature vector at each pixel position in the high-resolution continuous feature map to output a K-dimensional feature vector, where K is the number of discrete ink levels; S2.2: The input probability of the K-dimensional feature vector is converted into a discrete ink level probability distribution by the Softmax function, which represents the probability of each pixel selecting a different ink level; S2.3: Generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution; S3: Perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

2. The method of halftone image super-resolution reconstruction according to claim 1, characterized in that: In step S1.2, the method for spatial feature enhancement of the primary feature map using variable large kernel attention is as follows: Spatial feature enhancement of the primary feature map is performed using variable large kernel attention, wherein the variable large kernel attention function is: ， in, Let be a variable large kernel attention function, representing the intermediate feature map output by the variable large kernel attention function that enhances global contextual information. This represents a 1x1 point convolution. Represent a Depth-separable convolutions with variable kernel size. Represent a Depth-dilated convolution with variable kernel size. For variable kernel attention, the kernel size decreases along the network depth, and X is the input primary feature map.

3. The halftone image super-resolution reconstruction method according to claim 1, characterized in that: In step S1.3, the method for frequency domain feature decomposition of the intermediate feature map is as follows: perform dual-tree complex wavelet forward transform on the input intermediate feature map to decompose the spatial domain features into the frequency domain, resulting in one low-frequency approximate sub-band and six high-frequency detail sub-bands in different directions; the six high-frequency detail sub-bands in different directions correspond to ±15°, ±45°, and ±75° directions, respectively; the low-frequency approximate sub-band and the six high-frequency detail sub-bands constitute the multi-sub-band frequency domain features.

4. The halftone image super-resolution reconstruction method according to claim 1, characterized in that: In step S1.4, the method for feature fusion of the multi-subband frequency domain features by local linear multi-head self-attention is as follows: local linear multi-head self-attention divides the input multi-subband frequency domain features into multiple non-overlapping fixed-size windows in the spatial dimension; within each window, the multi-subband frequency domain features are aggregated and enhanced by the linear multi-head attention mechanism, and the output features of all windows are fused to generate deep enhanced features after spatial and frequency domain co-optimization.

5. The halftone image super-resolution reconstruction method according to claim 1, characterized in that: In step S2.3, the method for generating a preliminary halftone image of pixel-level discrete ink dot distribution based on the discrete ink level probability distribution is as follows: during the training phase, Gumbel-Softmax sampling is used to make differentiable discrete decisions on the discrete ink level probability distribution; during the inference phase, argmax operation is used to make hard decisions on the discrete ink level probability distribution to generate a preliminary halftone image of pixel-level discrete ink dot distribution.

6. The halftone image super-resolution reconstruction method according to claim 1, characterized in that: In step S3, the method for visual perception optimization of the preliminary halftone image is as follows: the preliminary halftone image is filtered in the frequency domain to suppress moiré noise at a specific frequency; an anisotropic diffusion algorithm is used to smooth artifacts in the image edge region and output a high-resolution halftone image.

7. A halftone image super-resolution reconstruction system, characterized in that, The modules are as follows: The super-resolution reconstruction module is used to extract and fuse multi-scale features from low-resolution continuous-tone images, and to improve spatial resolution to obtain high-resolution continuous feature maps. Specifically: S1.1: Perform initial feature mapping on the low-resolution continuous-tone image to extract a primary feature map containing basic features; S1.2: Spatial feature enhancement is performed on the primary feature map using variable large kernel attention to obtain an intermediate feature map with enhanced global context information; S1.3: Perform frequency domain feature decomposition on the intermediate feature map to obtain multi-subband frequency domain features; S1.4: The multi-subband frequency domain features are fused by local linear multi-head self-attention to obtain deep enhanced features that have been optimized by spatial and frequency domains. S1.5: Improve the spatial resolution of the deep enhancement features to obtain a high-resolution continuous feature map; The halftone generation module is used to discretize the high-resolution continuous feature map to obtain a discrete ink level probability distribution, and generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution, specifically: S2.1: Perform a linear transformation on the continuous feature vector at each pixel position in the high-resolution continuous feature map to output a K-dimensional feature vector, where K is the number of discrete ink levels; S2.2: The input probability of the K-dimensional feature vector is converted into a discrete ink level probability distribution by the Softmax function, which represents the probability of each pixel selecting a different ink level; S2.3: Generate a preliminary halftone image with pixel-level discrete ink dot distribution based on the discrete ink level probability distribution; The visual perception optimization module is used to perform visual perception optimization on the preliminary halftone image to obtain a high-resolution halftone image.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the halftone image super-resolution reconstruction method according to any one of claims 1 to 6.

Images