Infrared and visible image fusion method based on flow rate decoupling of W1 distance metric

Abstract

The present application belongs to the technical field of deep learning and computer vision, and discloses an infrared and visible light image fusion method based on W1 distance measurement flow speed decoupling. The present application solves the theoretical bottleneck that the nonlinear coupling in the high-dimensional feature space is difficult to completely eliminate by using a cross-modal image speed decoupling flow matching method based on the first-order Wassstein distance measurement. The present application can be used as a high-purity modal-agnostic feature extraction paradigm, providing strong bottom feature driving for complex remote sensing downstream tasks, and even achieving excellent performance on cross-modal data without field adaptation. The present application solves the difficulty that the traditional decoupling method is prone to numerical shock and feature collapse under small batch data training. The pseudo-distribution construction based on deterministic stagger and the minimax alternating optimization strategy not only realize the absolute decoupling of content and style at the dynamics level, but also generate cross-modal image data with high semantic consistency and pixel-level pairing.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and computer vision interdisciplinary technology, and relates to an infrared and visible light image fusion method based on W1 distance metric flow velocity decoupling. Background Technology

[0002] With the rapid development of artificial intelligence observation technology, the joint application of multimodal images has become a core means to improve the perception capability of complex scenes. Among the many modalities, infrared images (generated based on the thermal radiation characteristics of objects) and visible light images (generated based on the light reflection characteristics of object surfaces) constitute the most typical and complementary cross-modal data pair. In recent years, in order to provide sufficient virtual data support in downstream tasks (such as object detection and semantic segmentation), generative models have made great progress in the multimodal field. Early research relied on generative adversarial networks (GANs) or basic denoising diffusion probability models (DDPM), followed by the emergence of large pre-trained models specifically for infrared and visible images, such as MetaEarth, Geosynth, and DRS-Diff. Although these models perform well in single-modal generation, according to the latest review on infrared and visible light image fusion and conversion published in 2025, due to the lack of a strict feature separation mechanism, when dealing with complex and extreme scenes and cross-modal mapping, the geometric content and style attributes between different modalities are prone to data incompatibility and catastrophic entanglement.

[0003] To improve the stability and accuracy of cross-modal mapping, the academic community has begun to shift towards continuous flow models based on ordinary differential equations (ODEs). The introduction of flow matching technology has provided an efficient and simulation-free dynamic framework for generative modeling. In recent research, the theory of optimal transport (OT) has been deeply incorporated into flow matching networks to optimize generated trajectories. For example, Optimal Flow Matching, published at NeurIPS 2024, proposed a method to directly recover the optimal straight-line transport displacement in a single flow matching step; while the 2025 research, Beyond Optimal Transport: Model-Aligned Coupling, aims to further alleviate the path crossing problem through model alignment to generate high-quality samples. Furthermore, the latest framework FlowInOne in 2026 attempts to unify multi-channel control and heterogeneous modal mapping under a single image input-output flow matching architecture to address the inherent modal gap and marker artifact problems in diffusion models. However, while these cutting-edge advances optimize the geometric path of overall dynamic transport, they still fail to fundamentally solve the core problem of complete decoupling of nonlinear features.

[0004] Current cross-modal feature decoupling techniques suffer from a fundamental theoretical blind spot and technical bottleneck: traditional decoupling methods heavily rely on the assumption of geometric orthogonality, failing to achieve true statistical independence in high-dimensional manifold spaces. Specifically, existing decoupling paradigms, when constraining content features and style features to be mutually independent, generally employ dot product constraints (i.e., forcing the dot product to be zero) or cosine similarity penalties. However, in a strictly statistical sense, geometric perpendicularity is merely equivalent to the Pearson correlation coefficient of two random variables being zero; it only eliminates the linear correlation between variables. For high-dimensional, highly nonlinear latent space features extracted by deep neural networks, linear uncorrelation absolutely does not equate to information independence. Even if two feature vectors are strictly orthogonal in geometric space, they may still exhibit complex nonlinear coupling in high-dimensional manifolds (for example, small changes in infrared style features can still implicitly lead to distortions in the architectural outlines of content features).

[0005] To fundamentally sever this nonlinear feature dependency, we must evolve from geometrical orthogonality to statistical independence. In probability theory, the necessary and sufficient condition for two random variables to be completely decoupled is that their joint probability distribution must be strictly equal to the product of their marginal probability distributions. In generative models, traditional distance metrics (such as KL divergence) are prone to gradient vanishing or model collapse when dealing with non-overlapping high-dimensional manifolds. In contrast, the Wasserstein distance, as a metric mechanism based on the optimal transport problem, has been shown to provide a continuous and smooth gradient loss function when measuring the distribution of non-overlapping image manifolds, and due to its high degree of conformity to human visual perception, it is widely used in cutting-edge frameworks for adversarial generation and image mapping. Meanwhile, the sliced ​​Wasserstein distance, calculated by decomposing high-dimensional distributions into one-dimensional marginal distributions, has also become an effective means of approximating high-dimensional differences in generative models.

[0006] In summary, the current field of cross-modal feature decoupling urgently needs a new metric paradigm that can break through the limitations of linear orthogonality and provide a mechanism that can accurately measure and achieve absolute statistical independence in a continuous evolutionary trajectory, thereby realizing true and complete cross-modal decoupling of infrared and visible light. Summary of the Invention

[0007] The core objective of this invention is to address the industry-wide challenge of completely eliminating nonlinear coupling in high-dimensional feature spaces. In complex multimodal image generation and representation learning, content geometry and modal style often exhibit a deeply nonlinear interweaving, which traditional linear decoupling methods struggle to address. By innovatively constructing a pseudo-edge multiplicative distribution based on deterministic misalignment and deeply integrating it with a minimax alternating optimization adversarial strategy, this invention achieves a dynamic decoupling of feature distributions through a "dimensionality reduction attack" within the rigorous mathematical framework of the first-order Wasserstein-1 Distance. This mechanism completely removes redundant dependencies between modalities from the probability measure space, providing highly pure and robust modality-agnostic features for complex downstream computer vision tasks.

[0008] The technical solution of the present invention: An infrared-visible image fusion method based on W1 distance metric flow velocity decoupling, comprising the following steps: (1) Decomposition of total velocity field based on flow matching and constraint of anti-collapse kinetic energy loss: First, within the stream matching-based image generation network framework, the input for each training batch is set as... Latent variables of group-matched infrared and visible light images Latent variables Include Infrared latent variables and There are 1 pair of visible light latent variables, where It is a positive integer, either 3 or 4; within the current training batch, the total flow rate is decomposed into independent flow rate components using three parallel network branches: the latent variables of the current training batch... Parallel inputs are fed into the frozen pre-trained StableDiffusionXL model as the backbone network, the Content LoRA module for extracting content flow rate, and the Style LoRA module for extracting modal style. Latent variables are then processed through feature extraction and forward propagation operations. Perform mapping calculations and output the results to provide cross-domain general generative priors and basic evolutionary trajectories as global base velocities. Content flow rate is used to characterize modality-independent physical geometry, scene topology, and high-frequency edge contour features. and style flow rate used to characterize the radiative evolution characteristics, luminosity distribution, and thermodynamic texture features of the corresponding modes. The total flow rate is calculated through forward propagation of each network branch. for: To prevent content flow rate during the total flow rate breakdown process The physical geometry is implicitly encoded into style flow rate. This leads to the failure of feature decoupling. Therefore, a kinetic loss calculation node is set at the output of the style LoRA module to calculate the style flow rate output by the current training batch. of Norms, constructing constraints for kinetic energy loss to prevent collapse : (2) Decoupling and measurement of flow velocity characteristics based on probability distribution independence determination: To achieve style flow rate and content flow rate Decoupling and separation, proposing style flow rate and content flow rate The joint probability distribution is equal to the product of their respective marginal probability distributions, which serves as the criterion for complete decoupling. Statistical independence is determined on the probability manifold. Within the current training batch, the probability distribution extracted from each sample is used... Group content flow rate and Style flow rates are grouped, and joint probability distributions are constructed for each group. The sampling control group and the independent marginal multiplication distribution The sampling control group and the two groups are then sent together into the discriminator. The first-order Wassstein distance, which reflects the nonlinear dependence between the joint probability distribution and the multiplicative distribution of independent margins, is estimated by fitting. Step 1: Construct a joint probability distribution representing the current coupling state of image flow velocity features. The content flow rate and style flow rate pairs generated from the same source image within the current training batch and located at the same index position in the training batch are directly extracted to form a flow rate feature pair sampling set. ,in The algebraic index of the sample within the current training batch and ; The second step is to accurately simulate the distribution of independent edge multiplication integrals without violating the independence assumption. A deterministic cyclic translation operator for spatial state is introduced in the batch dimension of the current training batch. Perform a desort operation to force the current training batch to remove the first element from the desort list. The content flow rate and the first The mathematical expression for binding and combining style flow rates is as follows: In the formula, S is the deterministic cyclic translation operator of the spatial state; mod N is the mathematical modular operation; The third step is to construct the joint probability distribution. The sampling control group and the independent marginal multiplication distribution The sampled control group is fed into a discriminator composed of a multilayer fully connected MLP. Forward propagation calculations are performed to construct a first-order Wassean distance objective function for quantitatively measuring the nonlinear dependence between the joint probability distribution and the independent marginal product distribution. : In the formula, The first-order Wassstein distance; For mathematical expectation operators; To represent the velocity feature pairs from the joint probability distribution Obtained by sampling in the middle; To represent the velocity characteristic pairs as multiplicative distributions from independent edges Obtained by sampling in the middle; For discriminator The training parameters; Using the first-order Wassstein distance objective function Calculate the divergence difference between the joint probability distribution and the independent marginal multiplicative distribution.

[0009] (3) Minimax alternation adversarial optimization and bidirectional cross-modal exchange reconstruction: In order to ensure semantic consistency of images while completely separating content flow rate and style flow rate, an adversarial game mechanism and bidirectional cross-modal reconstruction constraints are introduced. For the sampling control group constructed in step (2), the mutual exclusion of flow rate in feature space and cross-modal reconstruction accuracy in image space are optimized in collaboration between the current training batch and the global level through alternating gradient updates of discriminator and generator, so as to achieve complete decoupling of image content and image style and global balance of reconstructed image quality.

[0010] Step 1: Discriminator parameter update phase: The constructed joint probability distribution The sampling control group and the independent marginal multiplication distribution The sampled control group is fed into the discriminator for forward propagation calculation, and a gradient penalty term is introduced into the objective function optimization of the discriminator. By forcibly satisfying the 1-Lipschitz continuity condition required for the dual form of the first-order Wassethan distance, backpropagation and gradient updates are then performed using the objective function of the constructed discriminator to maximize the value function. This maximization optimization operation searches for and quantifies the nonlinear coupling distance between content flow rate and style flow rate in the latent space of the current training batch, improving the discriminator's quantization accuracy of the probabilistic mutual information residue between the joint probability distribution and the independent marginal multiplication distribution; the discriminator's objective function The calculation formula is: In the formula, Represents the mathematical expectation. This represents the true joint distribution of content flow rate and style flow rate. Indicates the distribution of independent edge multiplication integrals. A discriminator indicating that the parameters are frozen; This represents the feature pair obtained by random interpolation between the true joint distribution and the pseudo-marginal multiplicative distribution. Represents the gradient operator. express Norm; Step 2: Generator Parameter Update Phase: After the discriminator completes its parameter update and provides adversarial gradients, the discriminator parameters are frozen. The content LoRA module and style LoRA module are then processed in the global joint loss function. Driven by this, adaptive gradient updates are performed; firstly, the content LoRA module and style LoRA module are minimized as much as possible during forward propagation. The first-order Wassstein distance in the loss function creates a spatial repulsion force within the velocity space, forcing the content velocity and style velocity to separate within the velocity space; its first-order Wassstein distance is an adversarial loss term. The specific calculation formula is as follows: Simultaneously, by further minimizing the constructed anti-collapse kinetic energy loss constraint Utilizing style flow rate of The norm imposes strong constraints on the output of the style LoRA module to suppress its ability to represent geometric structures and maintain the decoupling purity of the style flow rate. Finally, to ensure that the semantic consistency of the image is maintained after the content flow rate and style flow rate are separated, a cross-domain exchange flow operation is performed to construct cross-modal reconstruction constraints. The specific operation process is as follows: extract the infrared modal content flow rate features output from the content LoRA module in the current training batch. Visible modal style flow velocity characteristics output from the Style LoRA module It is then fed into the cross-combination node for cross-modal hard splicing and fusion, simultaneously incorporating the visible light modal content flow rate characteristics. Infrared modal style velocity characteristics Two sets of decoupled and reconstructed hybrid features are obtained by cross-combining corresponding matching. These features are then fed into an ODE based on flow matching for continuous time-step deduction and calculation, outputting the reconstructed target visible light image and target infrared image respectively. The reconstructed target visible light image and target infrared image are then compared with their corresponding real visible light image and real infrared image to calculate the exchange reconstruction loss. The exchange reconstruction loss term... The specific calculation formula is as follows: In the formula, and These are the reconstructed visible light image and the reconstructed infrared image of the target, respectively. and These are the corresponding real visible light image and real infrared image, respectively; The entire optimization process is ultimately determined by the global joint loss function. Unified Driver: .

[0011] The beneficial effects of this invention are as follows: This invention's cross-modal image velocity decoupling flow matching method based on the first-order Wassehstein distance metric solves the theoretical bottleneck of the difficulty in completely eliminating nonlinear coupling in high-dimensional feature spaces. This invention can serve as a high-purity modality-agnostic feature extraction paradigm, providing powerful underlying feature drivers for complex downstream remote sensing tasks (classification, segmentation, detection, etc.), and achieving excellent performance on cross-modal data even without domain adaptation. This invention addresses the difficulty of numerical oscillations and feature collapse easily generated by traditional decoupling methods under mini-batch data training. The proposed pseudo-distribution construction based on deterministic misplacement and the alternating Min-Max optimization strategy not only achieves absolute decoupling of content and style at the dynamic level but also generates semantically highly consistent and pixel-level paired cross-modal image data. Attached Figure Description

[0012] Figure 1 It is a stage architecture for velocity field dynamics decomposition and misaligned independence measurement.

[0013] Figure 2 It is a stage architecture of alternating minimax optimization and bidirectional cross-modal reconstruction. Detailed Implementation

[0014] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0015] Figure 1This is the architecture for the velocity field dynamics decomposition and misaligned independence measurement stage. First, iteratively samples image pairs of both visible (Vis) and infrared (IR) modes from the dataset, setting the training batch size to [value missing]. Then, the image pairs are mapped into the latent space to obtain latent variables that contain mixed feature representations of content and style. Next, the hidden variables... The data is fed in parallel into three independent velocity field dynamics decomposition modules: one part is fed into the Content LoRA module to extract cross-modal shared geometric structure and contour information to obtain the content flow rate. A portion of the feed is fed into the frozen pre-trained StableDiffusionXL model to maintain the basic generative dynamics and obtain the baseline flow rate. A portion is fed into the Style LoRA module to extract texture information such as photometric and thermal distribution for specific modalities, thus obtaining the style flow rate. The data obtained from the content LoRA module and style LoRA module will be processed by the distribution construction module, which first extracts data from the same image. and A strict one-to-one correspondence concatenation is performed at the batch level to obtain the joint probability distribution. Then, the style flow rate was... Introducing a core cyclic translation operator At the batch dimension, style features are physically misaligned (e.g., the content of the first sample is forcibly combined with the style of the second sample). After the misalignment operation, an independent edge multiplicative distribution is obtained. Then, these two distributions (the joint probability distribution and the independent marginal product distribution) are fed together into the Critic discriminator. Among them, the Wasserstein-1 (W) between them is calculated. Finally, by evaluating the divergence difference between the joint probability distribution and the independent edge multiplicative distribution, the network is quantitatively measured and optimized to achieve absolute mathematical independence between content features and style features.

[0016] Figure 2 It is a phased architecture of alternating minimax optimization and bidirectional cross-modal reconstruction. First, it will extract the content flow rate of the infrared modality output from the content LoRA module. and style flow rate of visible light modes output from the style LoRA module The extracted components are then fed into a cross-modal fusion node. The content flow rate of the visible light modality is then... Style flow rate with infrared mode Corresponding matching and cross-combination are also performed. Then, these two sets of "decoupled-recombined" hybrid features are fed into a flow-matching-based ODE for continuous time-step deduction and calculation, outputting the reconstructed target visible light image (Target Vis, ) and target infrared image (Target IR, The generated image and the real image are used to calculate the swap reconstruction loss. This forces the content LoRA module to retain only geometric structure, while the style LoRA module is only responsible for visual rendering. It's worth noting that throughout the feature extraction and generation process, to completely prevent the network from "hiding" structural information in the style flow, the output of the style LoRA module is specifically designed to... and Forced calculation of kinetic energy penalty loss This ensures that the energy complexity of style branches is strictly suppressed, much like a physical flow limiter. Finally, a min-max alternating game strategy governs the entire network: the discriminator is independently updated during the maximization phase to provide the most stringent... Distance measurement; during the minimization phase, by inputting the total joint loss after merging. (Include , and This guides the global parameter updates of the LoRA module. Through this strict closed-loop constraint, our network can generate multiple pixel-level aligned, content-style-completely decoupled multimodal paired images, providing high-quality data support for downstream tasks such as cross-modal object detection and semantic segmentation in a data-driven manner.

Claims

1. An infrared-visible image fusion method based on W1 distance metric flow velocity decoupling, characterized in that, The steps are as follows: (1) Decomposition of total velocity field based on flow matching and constraint of anti-collapse kinetic energy loss: First, within the stream matching-based image generation network framework, the input for each training batch is set as... Latent variables of group-matched infrared and visible light images Latent variables Include Infrared latent variables and There are 1 pair of visible light latent variables, where It is a positive integer, either 3 or 4; within the current training batch, the total flow rate is decomposed into independent flow rate components using three parallel network branches: the latent variables of the current training batch... Parallel inputs are fed into the frozen pre-trained StableDiffusionXL model as the backbone network, the Content LoRA module for extracting content flow rate, and the Style LoRA module for extracting modal style. Latent variables are then processed through feature extraction and forward propagation operations. Perform mapping calculations and output the results to provide cross-domain general generative priors and basic evolutionary trajectories as global base velocities. Content flow rate is used to characterize modality-independent physical geometry, scene topology, and high-frequency edge contour features. and style flow rate used to characterize the radiative evolution characteristics, luminosity distribution, and thermodynamic texture features of the corresponding modes. The total flow rate is calculated through forward propagation of each network branch. for: To prevent content flow rate during the total flow rate breakdown process The physical geometry is implicitly encoded into style flow rate. This leads to the failure of feature decoupling. Therefore, a kinetic loss calculation node is set at the output of the style LoRA module to calculate the style flow rate output by the current training batch. of Norms, constructing constraints for kinetic energy loss to prevent collapse : (2) Decoupling and measurement of flow velocity characteristics based on probability distribution independence determination: To achieve style flow rate and content flow rate Decoupling and separation, proposing style flow rate and content flow rate The joint probability distribution is equal to the product of their respective marginal probability distributions, which serves as the criterion for complete decoupling. Statistical independence is determined on the probability manifold. Within the current training batch, the probability distribution extracted from each sample is used... Group content flow rate and Style flow rates are grouped, and joint probability distributions are constructed for each group. The sampling control group and the independent marginal multiplication distribution The sampling control group and the two groups are then sent together into the discriminator. The first-order Wassstein distance, which reflects the nonlinear dependence between the joint probability distribution and the multiplicative distribution of independent margins, is estimated by fitting. (3) Minimax alternation adversarial optimization and bidirectional cross-modal exchange reconstruction: In order to ensure semantic consistency of images while completely separating content flow rate and style flow rate, an adversarial game mechanism and bidirectional cross-modal reconstruction constraints are introduced. For the sampling control group constructed in step (2), the mutual exclusion of flow rate in feature space and cross-modal reconstruction accuracy in image space are optimized in collaboration between the current training batch and the global level through alternating gradient updates of discriminator and generator, so as to achieve complete decoupling of image content and image style and global balance of reconstructed image quality.

2. The infrared-visible image fusion method based on W1 distance metric flow velocity decoupling according to claim 1, characterized in that, The specific implementation process of step (2) is as follows: Step 1: Construct a joint probability distribution representing the current coupling state of image flow velocity features. The content flow rate and style flow rate pairs generated from the same source image within the current training batch and located at the same index position in the training batch are directly extracted to form a flow rate feature pair sampling set. ,in The algebraic index of the sample within the current training batch and ; The second step is to accurately simulate the distribution of independent edge multiplication integrals without violating the independence assumption. A deterministic cyclic translation operator for spatial state is introduced in the batch dimension of the current training batch. Perform a desort operation to force the current training batch to remove the first element from the desort list. The content flow rate and the first The mathematical expression for binding and combining style flow rates is as follows: In the formula, S is the deterministic cyclic translation operator of the spatial state; mod N is the mathematical modular operation; The third step is to construct the joint probability distribution. The sampling control group and the independent marginal multiplication distribution The sampled control group is fed into a discriminator composed of a multilayer fully connected MLP. Forward propagation calculations are performed to construct a first-order Wassean distance objective function for quantitatively measuring the nonlinear dependence between the joint probability distribution and the independent marginal product distribution. : In the formula, The first-order Wassstein distance; For mathematical expectation operators; To represent the velocity feature pairs from the joint probability distribution Obtained by sampling in the middle; To represent the velocity characteristic pairs as multiplicative distributions from independent edges Obtained by sampling in the middle; For discriminator The training parameters; Using the first-order Wassstein distance objective function Calculate the divergence difference between the joint probability distribution and the independent marginal multiplicative distribution.

3. The infrared-visible image fusion method based on W1 distance metric flow velocity decoupling according to claim 1, characterized in that, The specific implementation process of step (3) is as follows: Step 1: Discriminator parameter update phase: The constructed joint probability distribution The sampling control group and the independent marginal multiplication distribution The sampled control group is fed into the discriminator for forward propagation calculation, and a gradient penalty term is introduced into the objective function optimization of the discriminator. By forcibly satisfying the 1-Lipschitz continuity condition required for the dual form of the first-order Wassethan distance, backpropagation and gradient updates are then performed using the objective function of the constructed discriminator to maximize the value function. This maximization optimization operation searches for and quantifies the nonlinear coupling distance between content flow rate and style flow rate in the latent space of the current training batch, improving the discriminator's quantization accuracy of the probabilistic mutual information residue between the joint probability distribution and the independent marginal multiplication distribution; the discriminator's objective function The calculation formula is: In the formula, Represents the mathematical expectation. This represents the true joint distribution of content flow rate and style flow rate. Indicates the distribution of independent edge multiplication integrals. A discriminator indicating that the parameters are frozen; This represents the feature pair obtained by random interpolation between the true joint distribution and the pseudo-marginal multiplicative distribution. Represents the gradient operator. express Norm; Step 2: Generator Parameter Update Phase: After the discriminator completes its parameter update and provides adversarial gradients, the discriminator parameters are frozen. The content LoRA module and style LoRA module are then processed in the global joint loss function. Driven by this, adaptive gradient updates are performed; firstly, the content LoRA module and style LoRA module are minimized as much as possible during forward propagation. The first-order Wassstein distance in the loss function creates a spatial repulsion force within the velocity space, forcing the content velocity and style velocity to separate within the velocity space; its first-order Wassstein distance is an adversarial loss term. The specific calculation formula is as follows: Simultaneously, by further minimizing the constructed anti-collapse kinetic energy loss constraint Utilizing style flow rate of The norm imposes strong constraints on the output of the style LoRA module to suppress its ability to represent geometric structures and maintain the decoupling purity of the style flow rate. Finally, to ensure that the semantic consistency of the image is maintained after the content flow rate and style flow rate are separated, a cross-domain exchange flow operation is performed to construct cross-modal reconstruction constraints. The specific operation process is as follows: extract the infrared modal content flow rate features output from the content LoRA module in the current training batch. Visible modal style flow velocity characteristics output from the Style LoRA module It is then fed into the cross-combination node for cross-modal hard splicing and fusion, simultaneously incorporating the visible light modal content flow rate characteristics. Infrared modal style velocity characteristics Two sets of decoupled and reconstructed hybrid features are obtained by cross-combining corresponding matching. These features are then fed into an ODE based on flow matching for continuous time-step deduction and calculation, outputting the reconstructed target visible light image and target infrared image respectively. The reconstructed target visible light image and target infrared image are then compared with their corresponding real visible light image and real infrared image to calculate the exchange reconstruction loss. The exchange reconstruction loss term... The specific calculation formula is as follows: In the formula, and These are the reconstructed visible light image and the reconstructed infrared image of the target, respectively. and These are the corresponding real visible light image and real infrared image, respectively; The entire optimization process is ultimately determined by the global joint loss function. Unified Driver: 。

Images