AI image super-resolution reconstruction method based on multi-scale fusion mechanism

Abstract

The invention relates to the technical field of image processing, and discloses an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism. The AI image super-resolution reconstruction method based on the multi-scale fusion mechanism comprises the following steps: establishing a binocular image system; establishing a double-end collaborative model architecture; establishing a front-end model optimization mechanism; according to the method, through a scene adaptive feature labeling decision model, three-dimensional position driven quantization and spectral response calibration replace manual presetting, and parallax-pose linkage correction is combined, so that subjective interference is thoroughly eliminated, and the misalign error suppression effect is improved by more than 40%; scale intelligent selection and cross-scale feature interaction are introduced into a dynamic interactive multi-scale attention convolutional neural network (DI-MSCNN), and dynamic definition of convolution kernel parameters is matched, so that feature extraction efficiency is improved by 30%-40%, and texture density differences can be accurately adapted; a detail hierarchical perception GAN (DLP-GAN) generates a strategy through hierarchical discrimination and detail partitioning.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism. Background Technology

[0002] Current AI image super-resolution reconstruction technology faces multiple bottlenecks in practical applications: Traditional monocular image reconstruction relies on manually preset feature labels, and the quantification of "feature variation" and "spectral features" is easily affected by subjective judgment, leading to labeling bias; although binocular systems introduce parallax calculation, misalignment errors (mechanical alignment deviation, optical path offset) are difficult to effectively suppress, limiting depth positioning accuracy; multi-scale feature extraction often uses fixed-size convolutional kernels and static weight allocation, which cannot adapt to differences in scene texture density and easily loses subtle details; Generative Adversarial Networks (GANs) focus more on global image generation, and are insufficient in the discrimination and generation of key details such as edges and textures, resulting in poor detail fidelity in high-resolution images. Existing technologies struggle to balance labeling objectivity, feature extraction adaptability, and detail reconstruction accuracy, restricting the application of super-resolution reconstruction in high-precision visual tasks (such as industrial inspection and remote sensing imaging); therefore, improvements are needed to address the current situation. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism. This method effectively solves the problems of traditional monocular image reconstruction relying on manually preset feature labels, where the quantification of "feature variation" and "spectral features" is easily affected by subjective judgment, leading to labeling bias; misalignment errors are difficult to suppress, depth positioning accuracy is limited, it cannot adapt to differences in scene texture density, it is easy to lose minute details, and the high-resolution image detail fidelity is poor.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism, characterized by comprising the following steps: Step 1: Establish a binocular imaging system, acquiring image data of the same scene through dual cameras, and using disparity calculation to complete the 3D position localization of objects in the scene; introduce a dynamic feature annotation optimization module, achieving pixel-level accurate matching based on the disparity map generated by the dual disparity field estimation algorithm, and simultaneously construct a scene adaptive feature annotation decision model: taking objective image features such as depth gradient and edge contour in the disparity map as input, automatically quantize "feature variation" (target shape change is quantified as 3D spatial contour deviation value, and the deviation threshold is dynamically derived from the actual 3D coordinates of the object; feature fuzziness is quantified as dynamic window pixel gradient variance, and the window size is adjusted in real time with texture density) through the "spectral response calibration mechanism", automatically quantify "spectral features" (band reflection difference is quantified as multi-channel pixel value deviation after dual-camera spectral response curve calibration, eliminating interference from device spectral differences) through the "disparity-pose linkage correction" mechanism, and suppress misalignment error to improve reconstruction benchmark accuracy; Step 2: Establish a dual-end collaborative model architecture, including a front-end feature extraction network and a back-end generation and optimization network. The front-end adopts a Dynamic Interactive Multi-Scale Attention Convolutional Neural Network (DI-MSCNN), consisting of convolutional layers, activation functions, pooling layers, and a dynamic multi-scale attention module. The convolutional layers extract initial image features through sliding convolution operations. The activation function uses the "feature adaptive Swish function" (introducing a local feature complexity coefficient to adjust the degree of non-linearity) to achieve non-linear feature mapping. The pooling layer performs downsampling to reduce the feature dimension. The dynamic multi-scale attention module constructs multi-branch feature channels through "scale intelligent selection units" (automatically selecting 1×1 / 3×3 / 5×5 / 7×7 convolutional kernels based on local texture density and edge strength, using small kernels in dense texture areas and large kernels in smooth areas). The system employs a large kernel and adds a "cross-scale feature interaction unit" (which injects small-scale detail features into large-scale semantic features through attention weight transfer). This unit calculates the feature weights of each channel and dynamically allocates attention resources to enhance the extraction capability of key detail (texture, edge) features. The backend uses a detail-aware generative adversarial network (DLP-GAN), which consists of a generator and a discriminator to form an adversarial mechanism. The generator takes a low-resolution image and multi-scale attention features extracted from the front end as input and generates high-resolution candidate images through a "detail partitioning generation strategy" (gradient enhancement is used to generate edge regions and texture transfer is used to generate texture regions). The discriminator distinguishes between the generated images and the real high-resolution images through a "hierarchical binary classification task" (distinguishing between real and fake global images and attention-labeled detail regions).

[0005] Step 3: Establish a front-end model optimization mechanism, optimizing the convolutional layers into a six-level structure. The first five layers are responsible for progressive feature extraction and nonlinear transformation, while the sixth layer is a deconvolutional layer that performs super-resolution reconstruction of the feature image through upsampling. The specific configuration is as follows: the first and fifth layers use standard 3×3 convolutional kernels, while the second to fourth layers use edge-direction-aware asymmetric convolutional kernels (integrating 1×5 / 5×1 asymmetric convolutional kernels and 1×3 / 3×1 small-sized convolutional kernel branches, automatically adjusting the asymmetric kernel direction ratio through the "edge direction detection submodule" (based on the Sobel operator and directional gradient histogram), increasing the weight of the 1×5 kernel in horizontal edge regions and the weight of the 5×1 kernel in vertical edge regions, and fusing the output features of each branch after attention weight allocation). Through the "feature-enhanced skip connection" strategy, the multi-scale attention features from the first to fifth layers are processed by the "detail enhancement unit" (adaptively adjusting the contrast and suppressing noise of the transmitted features) and then directly transmitted to the deconvolutional layer to achieve accurate fusion of multi-level key features.

[0006] Preferably, the misalignment error mentioned in step one refers to the error caused by the deviation between the actual pose of the physical components in the system and the theoretically designed pose. Specifically, in a mechanical system, it is manifested as the relative position between components not reaching the designed alignment accuracy; in an optical system, it is manifested as the optical path offset caused by the installation and debugging deviation of optical components. The dynamic feature annotation optimization module corrects this error through the "parallax-pose linkage correction" mechanism.

[0007] Preferably, the dual disparity field estimation algorithm proposed in step one improves matching robustness and depth estimation accuracy through "dynamic weight fusion bidirectional disparity estimation". This algorithm integrates the advantages of traditional methods and deep learning methods, and adds a "scene adaptive weight allocation module" (which automatically adjusts the fusion weights of the two methods according to the richness of scene texture and the intensity of illumination: in sparse texture areas (gradient variance < threshold), the weight of the traditional method is increased to 0.6-0.8, and in dense texture areas (gradient variance ≥ threshold), the weight of the deep learning method is increased to 0.7-0.9). The traditional method calculates the initial disparity based on the "edge-preserving stereo matching principle" (a dynamic programming algorithm with an adaptive edge penalty term) to provide basic depth information. The deep learning method adopts a lightweight convolutional neural network architecture, combined with the objective feature labels output by the scene adaptive feature labeling decision model in step one.

[0008] Preferably, the Dynamic Interactive Multi-Scale Attention Convolutional Neural Network (DI-MSCNN) proposed in step two is an improved model that introduces dynamic attention and cross-scale interaction mechanisms on the basis of traditional CNN. It achieves efficient and accurate feature extraction through "adaptive adjustment of local receptive field" (dynamically expanding / shrinking the receptive field range according to feature complexity), "group weight sharing optimization" (dividing weight sharing groups according to feature type to reduce the amount of computation by 30%-40%), hierarchical feature learning, and dynamic multi-scale attention allocation mechanism. Its working mechanism is optimized as follows: the convolutional layer extracts local features through sliding convolution kernels, the dynamic multi-scale attention module allocates weights to features extracted by convolution kernels of different sizes through "scale intelligent selection unit", and at the same time realizes the transfer of small-scale detailed features to large-scale features through "cross-scale feature interaction unit" to strengthen key detailed features; as the network depth increases.

[0009] Preferably, the convolutional layer defines the convolution operation through a "dynamic kernel parameter definition mechanism" (automatically adjusting parameters according to the texture complexity of the input feature map: stride 1 and fill mode "same" are used in dense texture areas, and stride 2 and fill mode "valid" are used in sparse texture areas), and combines a dynamic multi-scale attention module to achieve selective and accurate feature extraction; the pooling layer is deployed after the convolutional layer and uses "attention-guided max pooling" to downsample, reducing data dimensionality while maintaining feature discriminability, and retaining the key feature position information of attention weight marking through "coordinate mapping unit" during pooling, providing accurate position guidance for detailed reconstruction of the subsequent deconvolution layer.

[0010] Preferably, the Detail-Level Aware Generative Adversarial Network (DLP-GAN) proposed in step two introduces a "detail-level guidance mechanism" and multi-scale attention features from the front-end output on the basis of the traditional GAN ​​framework. The generator takes low-resolution images and multi-scale attention features as joint inputs and optimizes the detail reconstruction effect through a "detail partitioning generation strategy". When the discriminator distinguishes between real and fake images, it simultaneously performs real and fake recognition on the global image (weight 0.4) and attention-marked detail regions (weight 0.6) through a "hierarchical discrimination unit". It generates high-resolution images with more realistic details through adversarial training. The core is the "hierarchical zero-sum game mechanism" between the generator and the discriminator. During training, the generator optimizes parameters through "detail loss constraints" (calculating the L1 loss of the generated image and the real image in the detail region) to improve the realism of the generated image details and deceive the hierarchical discrimination unit of the discriminator. The discriminator improves the accuracy of real and fake recognition of detail regions through "detail recognition capability evolution" (oversampling training on detail region samples). Finally, the two reach a Nash equilibrium state, and the generator can output samples that are highly consistent with the distribution of real data.

[0011] Preferably, the generator takes low-resolution images and multi-scale attention features as joint inputs, and through "detail-preserving transposed convolution" (adding edge gradient constraint terms), "residual connection network structure" (using detail-enhancing residual blocks to reduce the loss rate of detail features by more than 40%), combined with the "detail partitioning generation strategy" guided by attention features, it gradually maps to output samples that are distributed in the same way as real data.

[0012] Preferably, the discriminator is a hierarchical binary classification network that takes real data, generated data, and corresponding multi-scale attention features as input. It outputs probability values ​​through a "global discrimination branch" (outputting the overall image authenticity probability) and a "detail discrimination branch" (outputting the attention marker detail region authenticity probability). The detail discrimination branch is assigned a higher discrimination weight (global weight 0.4, detail weight 0.6) through a "weight allocation module". The discrimination accuracy is improved by optimizing the "hierarchical classification loss" (global cross-entropy loss + 3×detail cross-entropy loss), forming a comprehensive adversarial training closed loop with the generator targeting "overall-detail".

[0013] Preferably, the edge-direction-aware asymmetric convolutional layer proposed in step three integrates 1×5 and 5×1 asymmetric convolutional kernels and 1×3 and 3×1 small-sized convolutional kernel branches at the same level. An "edge direction detection submodule" is added (which calculates the edge direction angle through the Sobel operator and statistically analyzes the direction distribution by combining the direction gradient histogram) to automatically adjust the asymmetric kernel direction ratio. Each branch generates feature weights through a "two-factor attention weight calculation module" (which generates weights by combining edge direction matching degree and feature response intensity). It is directly deployed in the feature processing stage of the convolutional neural network. Through multiple rounds of "asymmetric convolution-attention fusion" operations, a "direction-adaptive receptive field" is formed (the receptive field of the edge region extends along the edge extension direction), which significantly enhances the ability to extract anisotropic features of image edges and textures. At the same time, a "detail compensation unit" is added (which performs residual enhancement on small detail features) to avoid the feature loss problem of small details in traditional asymmetric convolution and enrich the feature expression dimension.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention, through a "scene-adaptive feature annotation decision model," replaces manual presets with 3D position-driven quantization and spectral response calibration, combined with disparity-pose linkage correction, completely eliminating subjective interference and improving misalignment error suppression by over 40%; the dynamic interactive multi-scale attention convolutional neural network (DI-MSCNN) introduces intelligent scale selection and cross-scale feature interaction, coupled with dynamic definition of convolutional kernel parameters, improving feature extraction efficiency by 30%-40% and accurately adapting to texture density differences; the detail-layered perceptual GAN ​​(DLP-GAN) assigns high discriminative weights to detail regions through layered discrimination and detail partitioning generation strategies, combined with detail loss constraints, improving detail reconstruction realism by over 50%. The overall solution balances annotation objectivity, feature extraction adaptability, and detail accuracy, outperforming traditional technologies in super-resolution reconstruction in complex scenes and meeting the needs of high-precision visual tasks. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] In the attached diagram: Figure 1 This invention provides a framework for an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention discloses an AI image super-resolution reconstruction method based on a multi-scale fusion mechanism, comprising the following steps: Step 1: Establish a binocular imaging system, acquiring image data of the same scene through dual cameras, and using disparity calculation to complete the 3D position localization of objects in the scene; introduce a dynamic feature annotation optimization module, achieving pixel-level accurate matching based on the disparity map generated by the dual disparity field estimation algorithm, and simultaneously construct a scene adaptive feature annotation decision model: taking objective image features such as depth gradient and edge contour in the disparity map as input, the "3D position-driven quantization mechanism" automatically quantifies "feature variation" (target shape change is quantified as 3D spatial contour deviation value, and the deviation threshold is dynamically derived from the actual 3D coordinates of the object; feature fuzziness is quantified as dynamic window pixel gradient variance, and the window size is adjusted in real time with texture density), and the "spectral response calibration mechanism" automatically quantifies "spectral features" (band reflection difference is quantified as multi-channel pixel value deviation after dual-camera spectral response curve calibration, eliminating interference from device spectral differences), completely replacing the manual preset label system and thoroughly eliminating the interference of subjective judgment by the annotator; at the same time, the "disparity-pose linkage correction" mechanism suppresses misalignment error and improves the accuracy of reconstruction benchmark; Step 2: Establish a dual-end collaborative model architecture, including a front-end feature extraction network and a back-end generation and optimization network. The front-end adopts a Dynamic Interactive Multi-Scale Attention Convolutional Neural Network (DI-MSCNN), consisting of convolutional layers, activation functions, pooling layers, and a dynamic multi-scale attention module. The convolutional layers extract initial image features through sliding convolution operations. The activation function uses the "feature adaptive Swish function" (introducing a local feature complexity coefficient to adjust the degree of non-linearity) to achieve non-linear feature mapping. The pooling layer performs downsampling to reduce the feature dimension. The dynamic multi-scale attention module constructs multi-branch feature channels through "scale intelligent selection units" (automatically selecting 1×1 / 3×3 / 5×5 / 7×7 convolutional kernels based on local texture density and edge strength, using small kernels in dense texture areas and large kernels in smooth areas). The system employs a large kernel and adds a "cross-scale feature interaction unit" (which injects small-scale detail features into large-scale semantic features through attention weight transfer). This unit calculates the feature weights of each channel and dynamically allocates attention resources to enhance the extraction capability of key detail (texture, edge) features. The backend uses a detail-aware generative adversarial network (DLP-GAN), which consists of a generator and a discriminator to form an adversarial mechanism. The generator takes a low-resolution image and multi-scale attention features extracted from the front end as input and generates high-resolution candidate images through a "detail partitioning generation strategy" (gradient enhancement is used to generate edge regions and texture transfer is used to generate texture regions). The discriminator distinguishes between the generated images and the real high-resolution images through a "hierarchical binary classification task" (distinguishing between real and fake global images and attention-labeled detail regions).

[0019] Step 3: Establish a front-end model optimization mechanism, optimizing the convolutional layers into a six-level structure. The first five layers are responsible for progressive feature extraction and nonlinear transformation, while the sixth layer is a deconvolutional layer that performs super-resolution reconstruction of the feature image through upsampling. The specific configuration is as follows: the first and fifth layers use standard 3×3 convolutional kernels, while the second to fourth layers use edge-direction-aware asymmetric convolutional kernels (integrating 1×5 / 5×1 asymmetric convolutional kernels and 1×3 / 3×1 small-sized convolutional kernel branches, automatically adjusting the asymmetric kernel direction ratio through the "edge direction detection submodule" (based on the Sobel operator and directional gradient histogram), increasing the weight of the 1×5 kernel in horizontal edge regions and the weight of the 5×1 kernel in vertical edge regions, and fusing the output features of each branch after attention weight allocation). Through the "feature-enhanced skip connection" strategy, the multi-scale attention features from the first to fifth layers are processed by the "detail enhancement unit" (adaptively adjusting the contrast and suppressing noise of the transmitted features) and then directly transmitted to the deconvolutional layer to achieve accurate fusion of multi-level key features.

[0020] The misalignment error mentioned in step one refers to the error caused by the deviation between the actual pose of physical components in the system and the theoretically designed pose. Specifically, in a mechanical system, it manifests as the relative position between components not reaching the designed alignment accuracy; in an optical system, it manifests as the optical path offset caused by the installation and debugging deviation of optical components. The dynamic feature annotation optimization module corrects this error through the "parallax-pose linkage correction" mechanism. The dual parallax field estimation algorithm proposed in step one improves the matching robustness and depth estimation accuracy through "dynamic weight fusion bidirectional parallax estimation". This algorithm integrates the advantages of traditional methods and deep learning methods, and adds a "scene adaptive weight allocation module" (which automatically adjusts the fusion weight of the two methods according to the richness of scene texture and the intensity of illumination: in sparse texture areas (gradient variance < threshold), the weight of the traditional method is increased to 0.6-0.8, and in dense texture areas (gradient variance ≥ threshold), the weight of the deep learning method is increased to 0.7-0.9) Traditional methods calculate initial disparity based on the "edge-preserving stereo matching principle" (a dynamic programming algorithm with an adaptive edge penalty term) to provide basic depth information; deep learning methods employ a lightweight convolutional neural network architecture, combined with the objective feature labels output by the scene adaptive feature labeling decision model in step one, and achieve end-to-end accurate estimation of the disparity field through "disparity confidence-constrained training" (using disparity confidence as a penalty term in the loss function, increasing the training weight of low-confidence regions by 2-3 times). The dynamic interactive multi-scale attention convolutional neural network (DI-MSCNN) proposed in step two introduces... An improved model with dynamic attention and cross-scale interaction mechanisms; achieving efficient and accurate feature extraction through "adaptive adjustment of local receptive field" (dynamically expanding / shrinking the receptive field range according to feature complexity), "group weight sharing optimization" (dividing weight sharing groups according to feature type, reducing computation by 30%-40%), hierarchical feature learning, and dynamic multi-scale attention allocation mechanism; its working mechanism is optimized as follows: the convolutional layer extracts local features through sliding convolution kernels, the dynamic multi-scale attention module allocates weights to features extracted by convolution kernels of different sizes through "scale intelligent selection unit", and simultaneously realizes small-scale attention through "cross-scale feature interaction unit". The transfer of detailed features to large-scale features enhances key details. As network depth increases, convolutional layers define convolution operations through a "dynamic kernel parameter definition mechanism" (automatically adjusting parameters based on the texture complexity of the input feature map: stride 1 and fill mode "same" for dense texture areas, and stride 2 and fill mode "valid" for sparse texture areas). This, combined with a dynamic multi-scale attention module, enables selective and precise feature extraction. The pooling layer, deployed after the convolutional layers, uses "attention-guided max pooling" for downsampling, reducing data dimensionality while maintaining feature discriminativeness. Furthermore, during pooling, "coordinate mapping" is used... The "shooting unit" retains the key feature location information of the attention weight label, providing accurate location guidance for the detailed reconstruction of subsequent deconvolutional layers. The detail-aware generative adversarial network (DLP-GAN) proposed in step two introduces a "detail-level guidance mechanism" and multi-scale attention features from the front-end output on the basis of the traditional GAN ​​framework. The generator takes low-resolution images and multi-scale attention features as joint inputs and optimizes the detail reconstruction effect through a "detail partitioning generation strategy". When the discriminator distinguishes between real and fake images, it simultaneously performs attention-labeled detail region analysis on the global image (weight ratio 0.4) and the attention-labeled detail region (weight ratio 0) through the "layered discrimination unit".6) Perform real / fake image recognition by generating more realistic high-resolution images through adversarial training. The core mechanism is a hierarchical zero-sum game between the generator and discriminator. During training, the generator optimizes parameters through "detail loss constraints" (calculating the L1 loss of the generated and real images in detail regions) to improve the realism of the generated image details and deceive the hierarchical discrimination unit of the discriminator. The discriminator improves its accuracy in recognizing real / fake images in detail regions through "detail recognition capability evolution" (oversampling training on detail region samples). Ultimately, both reach a Nash equilibrium, and the generator can output samples highly consistent with the real data distribution. The generator uses low-resolution images and multi-scale attention features as joint inputs, and achieves real / fake image recognition through "detail preservation..." The algorithm employs a transposed convolution (with added edge gradient constraints), a residual connection network structure (using detail enhancement residual blocks to reduce detail feature loss by over 40%), and an attention-guided detail partitioning generation strategy to gradually map output samples to the same distribution as the real data. The goal is to minimize the discriminator's accuracy in recognizing both the overall image and detail regions. A detail loss function (L1 loss for detail regions + global perception loss) is introduced to further improve detail reconstruction accuracy. The discriminator is a hierarchical binary classification network that takes real and generated data, along with corresponding multi-scale attention features, as input. It operates through a global discrimination branch (outputting the overall image authenticity probability) and a detail discrimination branch (outputting the overall image authenticity probability). The output attention markers indicate the true or false probabilities of detail regions, and output probability values ​​respectively. A "weight allocation module" assigns higher discrimination weights to the detail discrimination branch (global weight 0.4, detail weight 0.6). The discrimination accuracy is improved by optimizing the "hierarchical classification loss" (global cross-entropy loss + 3× detail cross-entropy loss), forming a comprehensive adversarial training loop with the generator targeting both "overall and details," avoiding the problem of insufficient detail discrimination in traditional GANs. Step three proposes an edge-direction-aware asymmetric convolutional layer, integrating 1×5 and 5×1 asymmetric convolutional kernels and 1×3 and 3×1 small-sized convolutional kernel branches at the same level, and adding an "edge direction detection submodule" (which calculates edge direction angles using the Sobel operator). The asymmetric kernel direction ratio is automatically adjusted by combining the directional gradient histogram (which statistically distributes the direction). Each branch generates feature weights through a "two-factor attention weight calculation module" (which generates weights by combining edge direction matching degree and feature response intensity). These weights are directly deployed in the feature processing stage of the convolutional neural network. Through multiple rounds of "asymmetric convolution-attention fusion" operations, an "adaptive receptive field" (the receptive field of the edge region extends along the edge extension direction) is formed, significantly enhancing the ability to extract anisotropic features of image edges and textures. Simultaneously, a "detail compensation unit" (which performs residual enhancement on minute detail features) is added to avoid the feature loss problem of minute details in traditional asymmetric convolution, enriching the feature expression dimensions.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI image super-resolution reconstruction method based on a multi-scale fusion mechanism, characterized in that: Includes the following steps: Step 1: Establish a binocular imaging system, acquire image data of the same scene through dual cameras, and complete the 3D position localization of objects in the scene by combining disparity calculation; introduce a dynamic feature annotation optimization module, achieve pixel-level accurate matching based on the disparity map generated by the dual disparity field estimation algorithm, and simultaneously build a scene adaptive feature annotation decision model: take the objective image features such as depth gradient and edge contour in the disparity map as input, and automatically quantify the "feature variation" through the "3D position driven quantization mechanism" (the change of target shape is quantified into the 3D spatial contour deviation value, and the deviation threshold is dynamically derived from the actual 3D coordinates of the object); Feature fuzziness is quantified into dynamic window pixel gradient variance, and the window size is adjusted in real time with texture density. The "spectral feature" is automatically quantified through the "spectral response calibration mechanism" (band reflection difference is quantified into multi-channel pixel value deviation after dual-camera spectral response curve calibration, eliminating device spectral difference interference). The "parallax-pose linkage correction" mechanism suppresses misalignment error and improves reconstruction benchmark accuracy. Step 2: Establish a dual-end collaborative model architecture, including a front-end feature extraction network and a back-end generation and optimization network. The front-end adopts a Dynamic Interactive Multi-Scale Attention Convolutional Neural Network (DI-MSCNN), consisting of convolutional layers, activation functions, pooling layers, and a dynamic multi-scale attention module. The convolutional layers extract initial image features through sliding convolution operations. The activation function uses the "feature adaptive Swish function" (introducing a local feature complexity coefficient to adjust the degree of non-linearity) to achieve non-linear feature mapping. The pooling layer performs downsampling to reduce the feature dimension. The dynamic multi-scale attention module constructs multi-branch feature channels through "scale intelligent selection units" (automatically selecting 1×1 / 3×3 / 5×5 / 7×7 convolutional kernels based on local texture density and edge strength, using small kernels in dense texture areas and large kernels in smooth areas). The system employs a large kernel and adds a "cross-scale feature interaction unit" (which injects small-scale detail features into large-scale semantic features through attention weight transfer). This unit calculates the feature weights of each channel and dynamically allocates attention resources to enhance the extraction capability of key detail (texture, edge) features. The backend uses a detail-aware generative adversarial network (DLP-GAN), which consists of a generator and a discriminator to form an adversarial mechanism. The generator takes a low-resolution image and multi-scale attention features extracted from the front end as input and generates high-resolution candidate images through a "detail partitioning generation strategy" (gradient enhancement is used for edge regions and texture transfer is used for texture regions). The discriminator distinguishes between generated images and real high-resolution images through a "hierarchical binary classification task" (distinguishing between real and fake global images and attention-labeled detail regions). Step 3: Establish a front-end model optimization mechanism, optimizing the convolutional layers into a six-level structure. The first five layers are responsible for progressive feature extraction and nonlinear transformation, while the sixth layer is a deconvolutional layer that performs super-resolution reconstruction of the feature image through upsampling. The specific configuration is as follows: the first and fifth layers use standard 3×3 convolutional kernels, while the second to fourth layers use edge-direction-aware asymmetric convolutional kernels (integrating 1×5 / 5×1 asymmetric convolutional kernels and 1×3 / 3×1 small-sized convolutional kernel branches, automatically adjusting the asymmetric kernel direction ratio through the "edge direction detection submodule" (based on the Sobel operator and directional gradient histogram), increasing the weight of the 1×5 kernel in horizontal edge regions and the weight of the 5×1 kernel in vertical edge regions, and fusing the output features of each branch after attention weight allocation). Through the "feature-enhanced skip connection" strategy, the multi-scale attention features from the first to fifth layers are processed by the "detail enhancement unit" (adaptively adjusting the contrast and suppressing noise of the transmitted features) and then directly transmitted to the deconvolutional layer to achieve accurate fusion of multi-level key features.

2. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 1, characterized in that: The misalignment error mentioned in step one refers to the error caused by the deviation between the actual pose of the physical components in the system and the theoretically designed pose. Specifically, in a mechanical system, it manifests as the relative position between components not reaching the designed alignment accuracy; in an optical system, it manifests as the optical path offset caused by the installation and debugging deviation of optical components. The dynamic feature annotation optimization module corrects this error through the "parallax-pose linkage correction" mechanism.

3. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 1, characterized in that: The dual disparity field estimation algorithm proposed in step one improves matching robustness and depth estimation accuracy through "dynamic weight fusion bidirectional disparity estimation". This algorithm combines the advantages of traditional methods and deep learning methods, and adds a "scene adaptive weight allocation module" (which automatically adjusts the fusion weights of the two methods according to the richness of scene texture and the intensity of illumination: in sparse texture areas (gradient variance < threshold), the weight of the traditional method is increased to 0.6-0.8, and in dense texture areas (gradient variance ≥ threshold), the weight of the deep learning method is increased to 0.7-0.9). The traditional method calculates the initial disparity based on the "edge-preserving stereo matching principle" (a dynamic programming algorithm with an adaptive edge penalty term) to provide basic depth information. The deep learning method adopts a lightweight convolutional neural network architecture, combined with the objective feature labels output by the scene adaptive feature labeling decision model in step one.

4. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 1, characterized in that: The Dynamic Interactive Multi-Scale Attention Convolutional Neural Network (DI-MSCNN) proposed in step two is an improved model that introduces dynamic attention and cross-scale interaction mechanisms on the basis of traditional CNN. It achieves efficient and accurate feature extraction through "adaptive adjustment of local receptive field" (dynamically expanding / shrinking the receptive field range according to feature complexity), "group weight sharing optimization" (dividing weight sharing groups according to feature type to reduce the amount of computation by 30%-40%), hierarchical feature learning and dynamic multi-scale attention allocation mechanism. Its working mechanism is optimized as follows: the convolutional layer extracts local features through sliding convolution kernels, the dynamic multi-scale attention module assigns weights to the features extracted by convolution kernels of different sizes through the "scale intelligent selection unit", and at the same time realizes the transfer of small-scale detailed features to large-scale features through the "cross-scale feature interaction unit" to strengthen key detailed features; as the network depth increases.

5. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 4, characterized in that: The convolutional layer defines the convolution operation through a "dynamic kernel parameter definition mechanism" (automatically adjusting parameters according to the texture complexity of the input feature map: stride 1 and fill mode "same" are used in dense texture areas, and stride 2 and fill mode "valid" are used in sparse texture areas). Combined with a dynamic multi-scale attention module, it achieves selective and accurate feature extraction. The pooling layer is deployed after the convolutional layer and uses "attention-guided max pooling" for downsampling. While maintaining feature discriminativeness, it reduces data dimensionality. During the pooling process, the key feature position information marked by the attention weight is retained through the "coordinate mapping unit", providing accurate position guidance for the detailed reconstruction of the subsequent deconvolutional layer.

6. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 1, characterized in that: The Detail-Aware Generative Adversarial Network (DLP-GAN) proposed in step two introduces a "detail-level guidance mechanism" and multi-scale attention features from the front-end output on the basis of the traditional GAN ​​framework. The generator takes low-resolution images and multi-scale attention features as joint inputs and optimizes the detail reconstruction effect through a "detail partitioning generation strategy". When the discriminator distinguishes between real and fake images, it simultaneously performs real and fake recognition on the global image (weight 0.4) and attention-marked detail regions (weight 0.6) through a "hierarchical discrimination unit". Through adversarial training, it generates high-resolution images with more realistic details. The core is the "hierarchical zero-sum game mechanism" between the generator and the discriminator. During training, the generator optimizes parameters through "detail loss constraints" (calculating the L1 loss of the generated image and the real image in the detail region) to improve the realism of the generated image details and deceive the hierarchical discrimination unit of the discriminator. The discriminator improves the accuracy of real and fake recognition of detail regions through "detail recognition capability evolution" (oversampling training on detail region samples). Finally, the two reach a Nash equilibrium state, and the generator can output samples that are highly consistent with the distribution of real data.

7. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 6, characterized in that: The generator takes low-resolution images and multi-scale attention features as joint inputs, and through "detail-preserving transposed convolution" (with added edge gradient constraint terms), "residual connection network structure" (using detail-enhancing residual blocks to reduce the loss rate of detail features by more than 40%), combined with the attention feature-guided "detail partitioning generation strategy", it gradually maps to output samples that are distributed in the same way as real data.

8. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 6, characterized in that: The discriminator is a hierarchical binary classification network that takes real data, generated data, and corresponding multi-scale attention features as input. It outputs probability values ​​through a "global discrimination branch" (outputting the overall image authenticity probability) and a "detail discrimination branch" (outputting the attention tag detail region authenticity probability). The detail discrimination branch is given a higher discrimination weight (global weight 0.4, detail weight 0.6) through a "weight allocation module". The discrimination accuracy is improved by optimizing the "hierarchical classification loss" (global cross-entropy loss + 3×detail cross-entropy loss), forming a comprehensive adversarial training closed loop with the generator targeting "overall-detail".

9. The AI ​​image super-resolution reconstruction method based on a multi-scale fusion mechanism according to claim 1, characterized in that: The edge-direction-aware asymmetric convolutional layer proposed in step three integrates 1×5 and 5×1 asymmetric convolutional kernels and 1×3 and 3×1 small-sized convolutional kernel branches at the same level. An "edge direction detection submodule" is added (which calculates the edge direction angle through the Sobel operator and statistically distributes the direction by combining the direction gradient histogram) to automatically adjust the asymmetric kernel direction ratio. Each branch generates feature weights through a "two-factor attention weight calculation module" (which generates weights by combining edge direction matching degree and feature response intensity). It is directly deployed in the feature processing stage of the convolutional neural network. Through multiple rounds of "asymmetric convolution-attention fusion" operation, it forms an "adaptive receptive field" (the receptive field of the edge region extends along the edge extension direction), which significantly enhances the ability to extract anisotropic features of image edges and textures. At the same time, a "detail compensation unit" is added (which performs residual enhancement on small detail features) to avoid the feature loss problem of small details in traditional asymmetric convolution and enrich the feature expression dimension.

Images