An image computation-based explainable artificial intelligence analysis and inference system

Abstract

The application belongs to the field of artificial intelligence, and particularly relates to an explainable artificial intelligence analysis and reasoning system based on image calculation, aiming to solve the problems of AI image decision process opacity and causality logic untraceability. The system comprises multi-scale feature extraction, semantic concept generation, causality reasoning engine, counterfactual explanation generation and man-machine collaborative verification interface, and realizes high-trust image analysis by constructing an explainable path from pixels to semantic concepts and then to causality chain.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an interpretable artificial intelligence analysis and reasoning system based on image computing. Background Technology

[0002] Artificial intelligence (AI) technology has made significant progress in image analysis and reasoning in recent years, demonstrating enormous potential, especially in high-risk applications such as computer vision, medical image diagnosis, and autonomous driving. However, as the complexity of deep learning models continues to increase, their decision-making processes are increasingly exhibiting "black box" characteristics, leading to a lack of transparency and traceability in model outputs. This severely restricts their deployment and application in environments requiring high trust and strong oversight. Explainable AI (XAI), as a key research direction to address this problem, aims to reveal the internal decision-making logic of models, enhancing users' trust in and controllability of system outputs.

[0003] Among these, image-based computational interpretability analysis methods focus on analyzing the model's dependence on specific regions or features in the input image through visualization, feature attribution, or counterfactual reasoning. These methods typically combine gradient information, attention mechanisms, or generative modeling to attempt to establish semantic relationships between input pixels and model predictions, thereby providing human users with intuitive and understandable decision-making support.

[0004] Existing technologies still face multiple challenges in achieving image-driven interpretable reasoning: mainstream attribution methods (such as Grad-CAM and LIME) often provide only local or approximate explanations, making it difficult to guarantee causal consistency and cross-sample stability; while attention mechanisms can highlight key regions, their weight allocation lacks strict semantic alignment and is susceptible to noise interference; furthermore, existing systems generally lack the ability to combine explanation results with high-level logical reasoning, failing to support explanation-based causal inference or hypothesis testing. In scenarios with extremely high requirements for decision reliability, such as medical diagnosis and industrial quality inspection, these shortcomings may lead to increased risk of misjudgment, low efficiency of human-machine collaboration, and even security and compliance issues. Therefore, there is an urgent need for a new type of artificial intelligence analysis system that can deeply integrate image computing and interpretable reasoning, possessing accuracy, robustness, and logical consistency. Summary of the Invention

[0005] The purpose of this invention is to provide an interpretable artificial intelligence analysis and reasoning system based on image computing to solve the "black box" problem that is common in current artificial intelligence systems in image understanding tasks. That is, the model decision-making process lacks transparency, the causal logic is untraceable, and the reasoning basis is difficult for humans to understand, thus limiting its reliable deployment and compliant application in scenarios with high reliability requirements such as medical diagnosis, industrial quality inspection, and autonomous driving.

[0006] The technical solution of this invention is an interpretable artificial intelligence analysis and reasoning system based on image computing, comprising an image input module, a multi-scale feature extraction module, a semantic concept generation module, a causal reasoning engine, a counterfactual explanation generation module, and a human-machine collaborative verification interface. The image input module receives raw image data and performs standardized preprocessing on it. The multi-scale feature extraction module, connected to the image input module, extracts local detail features, mesoscopic structural features, and global semantic features of the image through parallel convolutional neural network branches, and fuses these three types of features in the channel dimension to form a unified multi-scale feature tensor. The semantic concept generation module, connected to the multi-scale feature extraction module, maps the multi-scale feature tensor to a set of predefined visual concept spaces with human-understandable semantics. The visual concept space is constructed from ontology knowledge bases in fields such as medical anatomy terminology, industrial defect types, and traffic scene elements. Each visual concept corresponds to an interpretable semantic label and its spatial location region. The causal reasoning engine is connected to the semantic concept generation module. The concept generation module, based on a Bayesian causal graph model, models the causal dependencies between activated visual concepts, identifies key concept nodes that have a direct causal effect on the final decision result, and excludes interfering concepts that are only correlated but have no causal effect. The counterfactual interpretation generation module, connected to the causal inference engine, generates a counterfactual image variant with minimal perturbation, given the original image and the current decision result, such that the system outputs the opposite decision result under this variant. The minimal perturbation is limited to the spatial region corresponding to the key concept nodes identified by the causal inference engine and is achieved through a gradient-guided pixel-level optimization algorithm. The human-machine collaborative verification interface, connected to the counterfactual interpretation generation module and the semantic concept generation module, synchronously presents the original image, activated semantic concept regions, causal dependency graph, and counterfactual image variant in a visual form to domain experts and receives feedback signals from experts on the rationality of the explanation. The feedback signals are used to dynamically adjust the weight distribution of the visual concept space and the edge strength parameters of the causal graph to achieve continuous optimization of the system's interpretability.

[0007] Furthermore, the multi-scale feature extraction module includes three parallel residual network branches, configured as shallow, medium, and deep networks, respectively. The shallow network contains two residual blocks for capturing high-frequency local features such as edges and textures; the medium network contains four residual blocks for extracting component-level structural features; and the deep network contains six residual blocks for encoding scene-level global semantics. The output feature maps of the three branches are upsampled or downsampled to a uniform spatial resolution, then concatenated along the channel dimension, and channel compression and feature fusion are performed through a 1×1 convolutional layer to output a fused feature tensor with a dimension of 128.

[0008] Furthermore, the semantic concept generation module adopts a concept bottleneck network architecture, which includes a fully connected layer and a concept activation function. The fully connected layer maps the fused feature tensor to an N-dimensional concept vector, where N is the total number of predefined visual concepts. The concept activation function is a piecewise linear threshold function. When a component in the concept vector exceeds a preset activation threshold of 0.7, the corresponding semantic concept is determined to be in an active state. The contribution region of the concept in the original image is located through backpropagation. The contribution region is generated by a gradient-weighted class activation mapping method, with a spatial resolution of 1 / 4 of the input image.

[0009] Furthermore, the Bayesian causal graph constructed by the causal inference engine uses activated semantic concepts as nodes, and the causal edge weights between nodes are obtained through structural equation modeling. The structural equation model is trained based on historical decision datasets and uses maximum likelihood estimation to fit the response function of each concept node to its parent node. During the inference phase, the causal inference engine calculates the average causal effect value of each activated concept on the final decision variable. If the value is greater than the preset causal significance threshold of 0.3, the concept is marked as a key causal concept.

[0010] Furthermore, the pixel-level optimization process performed by the counterfactual explanation generation module is as follows: using the original image as the initial value, a differentiable perturbation is applied to the spatial region corresponding to the key causal concept; the loss function is defined as the weighted sum of decision confidence loss and image fidelity loss, wherein the decision confidence loss prompts the model to output a high confidence of the opposite category for the counterfactual image, and the image fidelity loss uses the L2 norm to constrain the perturbation amplitude to not exceed 10% of the original pixel value; the perturbation amount is iteratively updated through the Adam optimizer until the decision result is flipped and the perturbation amplitude meets the constraint condition.

[0011] Furthermore, the human-machine collaborative verification interface supports experts to label any causal edge in the causal graph with three categories: "accept", "reject", or "uncertain". The system automatically triggers a parameter fine-tuning mechanism based on the expert labels: if a causal edge is rejected multiple times, its edge weight in subsequent inference is reduced to 50% of the initial value; if it is accepted multiple times, its edge weight is increased to 150% of the initial value. The weight adjustment updates the internal parameters of the causal inference engine in real time through online learning.

[0012] Furthermore, the visual concept space supports dynamic expansion; when the human-machine collaborative verification interface receives a new semantic concept definition and its example image submitted by an expert, the system automatically incorporates it into the concept ontology library and updates the mapping parameters of the semantic concept generation module through an incremental learning mechanism, without having to retrain the entire system.

[0013] Furthermore, the system operates in a dual-channel reasoning mode; in the normal mode, the system only outputs the final decision result; in the interpretable mode, the system simultaneously outputs the decision result, the list of activated semantic concepts, the causal dependency graph, and the counterfactual image variant; the activation of the interpretable mode is dynamically triggered by the external security policy controller according to the risk level of the application scenario.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0015] This invention fundamentally overcomes the black-box limitations of traditional deep learning models by constructing a complete interpretable path from raw pixels to human-understandable semantic concepts, and then to a causal logic chain. The combination of multi-scale feature extraction and conceptual bottleneck networks ensures that the system can both perceive image details and express them in structured semantics. The introduction of a causal inference engine enables the system to distinguish between correlation and causality, accurately identify key factors influencing decision-making, and avoid being misled by irrelevant features. The counterfactual explanation generation mechanism provides intuitive and verifiable evidence for decision comparison, greatly enhancing human trust in AI judgments. The human-machine collaborative verification interface realizes a closed-loop feedback between expert knowledge and model reasoning, enabling the system's interpretability to continuously evolve. The overall architecture balances high-performance image analysis capabilities with highly transparent decision-making logic, providing solid technical support for the safe and reliable application of artificial intelligence in high-risk and high-responsibility fields. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0017] Figure 2 This is a schematic diagram illustrating the core principle framework of the collaborative work between the causal reasoning engine and the counterfactual explanation generation module in this invention. Detailed Implementation

[0018] Example 1

[0019] Please refer to Figure 1 and Figure 2 This invention provides an interpretable artificial intelligence analysis and reasoning system based on image computing. Its core objective is to address the "black box" problem prevalent in existing image understanding AI systems, namely, the lack of transparency in the model's decision-making process, the untraceability of causal logic, and the difficulty for humans to understand the reasoning basis. By constructing a complete interpretable path from raw image pixel data to human-understandable semantic concepts, and then to the causal logic chain, this system fundamentally overcomes the limitations of traditional deep learning models in image understanding tasks.

[0020] This system architecture comprises an image input module, a multi-scale feature extraction module, a semantic concept generation module, a causal inference engine, a counterfactual explanation generation module, and a human-machine collaborative verification interface. The overall technical flow begins with the image input module receiving and preprocessing raw image data. Subsequently, the multi-scale feature extraction module extracts multi-dimensional features from the image. The extracted features are then sent to the semantic concept generation module, which maps them to visual concepts with human-understandable semantics. These activated visual concepts and their spatial location information are then used by the causal inference engine to construct causal dependencies and identify key concept nodes that have a direct causal effect on the final decision outcome. To verify and explain the decision, the counterfactual explanation generation module generates counterfactual image variants based on the key causal concepts. Finally, the human-machine collaborative verification interface simultaneously presents the original image, semantic concepts, causal dependency graph, and counterfactual image variants to domain experts and receives feedback signals to continuously optimize the system's interpretability. This system also supports a dual-channel inference mode, dynamically switching between a conventional mode and an interpretable mode according to the risk level of the application scenario, and the visual concept space has dynamic expansion capabilities.

[0021] The image input module aims to receive raw image data from various sources and perform necessary standardized preprocessing to ensure that subsequent processing units can stably and efficiently utilize the data. This module is an integrated data access and preprocessing platform that supports input of various image data formats, such as JPEG, PNG, DICOM medical image formats, and raw pixel streams acquired by various sensors. When the module receives raw image data, it initiates a series of automated preprocessing steps. The first step is image format parsing and decoding, converting the raw binary data into a workable pixel matrix. Next, data integrity verification is performed, such as checking whether the file header and image size meet expectations, and whether there is any data corruption or missing data; if an anomaly is detected, the system generates an error report and attempts data recovery or anomaly marking. Then, denoising operations are performed, such as using Gaussian filtering or median filtering algorithms, to smooth the image and reduce noise interference, which is especially crucial for sensor-acquired images or low-quality images. Finally, brightness and contrast adjustments are performed, such as using histogram equalization techniques to enhance the visual recognizability and feature distinguishability of the image. Size normalization is a crucial step. The module uniformly scales all images to a preset standard spatial resolution, such as 256 pixels by 256 pixels or 512 pixels by 512 pixels. Common interpolation algorithms include bilinear interpolation or bicubic interpolation to ensure consistent input size for the feature extraction module. Pixel value normalization maps pixel intensity values ​​to a standard range, such as between 0 and 1, or performs zero-mean unit variance normalization. This helps stabilize the training process of the neural network and improves convergence speed. For specific applications, such as medical diagnosis, color space conversion may also be performed, such as converting RGB images to grayscale or Lab color space, to highlight specific information or reduce redundancy. All these preprocessing steps are executed according to a predefined processing pipeline, ensuring that the output image data is a uniformly formatted, standardized, and reliable tensor, ready to be input into the multi-scale feature extraction module. In terms of anomaly handling, if an image is severely damaged beyond repair due to external factors, the module will trigger a rejection mechanism and return an error code and log information to the next higher-level caller to prevent contamination of subsequent processing flows.

[0022] The multi-scale feature extraction module, connected to the image input module, aims to capture local detail features, mesoscopic structural features, and global semantic features of an image from different levels of abstraction through parallel convolutional neural network branches. It then efficiently fuses these heterogeneous features along the channel dimension to form a unified and informative multi-scale feature tensor. The module's architecture is designed to simulate the human eye's cognitive process of observing objects, focusing on both subtle details and overall contours. The module comprises three parallel residual network branches, configured as shallow, mid-level, and deep networks, each focusing on extracting features at a specific scale.

[0023] Shallow networks consist of two residual blocks, whose primary task is to capture high-frequency local features such as edges and textures in an image. Each residual block typically includes two convolutional layers, two batch normalization layers, and two activation function layers, employing a skip connection mechanism to directly add the input to the block's output. This mitigates the vanishing gradient problem during deep network training and enhances feature representation capabilities. The convolutional kernels in shallow networks are typically small, such as 3x3, designed to accurately capture pixel gradient changes in local image regions, such as detecting object contours and texture details. The network outputs feature maps with high spatial resolution and a small number of channels, directly reflecting the original visual elements of the image. As data flows through the shallow network, each convolutional operation extracts edge information with different directions and frequencies. The batch normalization layers stabilize the distribution of these features, while activation functions, such as rectified linear units, introduce non-linear transformations, allowing them to be combined into more complex local patterns.

[0024] The mid-layer network contains four residual blocks, primarily used to extract component-level structural features. Compared to shallow networks, mid-layer networks have a larger receptive field, enabling them to aggregate information from multiple local features to form an abstract representation of object components in an image. For example, in a face image, a mid-layer network might identify features such as eyes, nose, and mouth; in industrial quality inspection scenarios, it might identify the shape of screws and the direction of weld seams. By increasing the number of residual blocks, mid-layer networks can perform deeper nonlinear transformations and feature abstraction. Their internal downsampling operations, such as using convolutional layers with a stride of 2, gradually reduce the spatial resolution of the feature map while increasing the number of channels, thereby encoding information at a higher semantic level.

[0025] Deep networks consist of six residual blocks, and their core function is to encode scene-level global semantics. Deep networks have the largest receptive field, enabling them to understand the overall content of an image, background information, and the spatial relationships between different components. For example, in autonomous driving scenarios, deep networks can identify "highway scenes" and "urban road scenes"; in medical images, they can identify the overall distribution of "tumor regions" and "healthy tissue regions." By further stacking residual blocks and downsampling, deep networks highly compress image information into low spatial resolution, high-channel-count feature maps. These feature maps contain the highest-level abstract semantic information of the image.

[0026] The output feature maps of the three parallel branches may differ in spatial resolution, thus requiring upsampling or downsampling operations, such as bilinear interpolation or deconvolution layers, to uniformly adjust them to a preset standard spatial resolution. After resolution unification, these feature maps are concatenated along the channel dimension. The concatenated feature tensor contains rich information from different scales and levels of abstraction. Finally, the concatenated tensor undergoes channel compression and feature fusion through a 1x1 convolutional layer. This 1x1 convolutional layer reduces dimensionality and further mixes information from different scales, while introducing nonlinear transformations to enhance feature representation capabilities, ultimately outputting a 128-dimensional fused feature tensor. This fused feature tensor is a high-dimensional vector that integrates local details, mesoscopic structure, and global semantic information of the image, providing comprehensive input for subsequent semantic concept generation. Regarding parameter management, the weights of each residual network branch can be pre-trained on large-scale image datasets, such as ImageNet, to obtain good general feature extraction capabilities, and then fine-tuned on specific task datasets to adapt to the specific application requirements of this system. The weight update uses the standard gradient descent optimization algorithm, combined with techniques such as batch normalization and weight decay, to ensure the stability of training and the generalization ability of the model.

[0027] The semantic concept generation module, connected to the multi-scale feature extraction module, aims to map the fused feature tensor output by the multi-scale feature extraction module to a predefined visual concept space with human-understandable semantics, and generate a corresponding semantic label and its spatial location region in the original image for each activated visual concept. The core design of this module is to transform the abstract features learned by the neural network into high-level semantic concepts that humans can intuitively understand, thereby achieving model interpretability.

[0028] This module employs a concept bottleneck network architecture. This architecture first includes a fully connected layer, which takes a 128-dimensional fusion feature tensor as input. Before entering the fully connected layer, the fusion feature tensor typically undergoes a global average pooling operation, compressing its spatial dimensions into a single vector, thus effectively transmitting global information from multi-scale features to the fully connected layer. The fully connected layer maps this pooled vector to an N-dimensional concept vector, where N is the total number of predefined visual concepts in the system. Each component corresponds to a specific visual concept, and its value represents the intensity or confidence of the concept's activation.

[0029] After generating the N-dimensional concept vector, the system applies a concept activation function. This activation function is a piecewise linear threshold function. Specifically, when the value of a component in the concept vector exceeds a preset activation threshold of 0.7, the corresponding semantic concept is considered active. This threshold setting aims to filter out low-confidence concepts, retaining only those semantic concepts with strong supporting evidence. For example, if a component of the concept vector represents "lung nodule" and its value reaches 0.8, then the semantic concept "lung nodule" is considered active. The activation function can be defined as: if x > 0.7, output x; otherwise, output 0, where x is a component of the concept vector. This piecewise linear function effectively achieves binary activation of concepts while preserving the gradient.

[0030] Once a semantic concept is determined to be active, the system uses backpropagation to locate the region in the original image that contributes to that concept's activation. Specifically, it employs Grad-weighted Activation Mapping (Grad-CAM). Grad-CAM calculates the gradient of the target concept on the last convolutional feature map and performs global average pooling on these gradients to generate a weight vector. This weight vector is then weighted and summed with the last convolutional feature map and processed by an activation function to generate a heatmap. This heatmap clearly indicates the region in the original image that contributes the most to activating the specific concept. The spatial localization region generated by Grad-CAM has a resolution of 1 / 4 of the input image, meaning it provides relatively fine-grained region localization information, but not at the pixel level. For example, for a 256-pixel by 256-pixel input image, the resolution of the localized region is 64-pixel by 64-pixel. This region can be a rectangular bounding box or a pixel-level mask, depending on the accuracy requirements of subsequent analysis.

[0031] The visual concept space is constructed from a predefined domain ontology knowledge base. This knowledge base is a structured knowledge system that contains all relevant concepts within a specific application domain, the hierarchical relationships between concepts, and their attributes. For example, in the field of medical diagnostics, the knowledge base might contain "medical anatomical terms" (such as "lung," "heart," "liver"), "disease types" (such as "pneumonia," "tumor"), and "imaging features" (such as "ground-glass opacity," "cavitation"). In the field of industrial quality inspection, it might contain "industrial defect types" (such as "scratches," "cracks," "foreign objects") and "part names" (such as "bolts," "welds"). In the field of traffic scene elements, it might contain "vehicle types," "pedestrians," "traffic signs," "lane lines," etc. These concepts and their semantic labels are defined and labeled by domain experts and managed and maintained through ontological tools to ensure their comprehensibility and standardization at the human cognitive level. Each visual concept not only has one semantic label but may also be associated with metadata describing its characteristics, such as typical shape and size range. When new information is collected, the system will pass the corresponding semantic tags and spatial location regions to the causal reasoning engine as the basic input for it to model causal relationships.

[0032] The causal reasoning engine, connected to the semantic concept generation module, aims to model the causal dependencies between visual concepts activated by the semantic concept generation module based on a Bayesian causal graph model. The core function of this engine is to identify key concept nodes that have a direct causal effect on the final decision outcome and to exclude interfering concepts that are only relevant but have no causal effect, thereby providing a more accurate and reliable explanation for the decision.

[0033] The Bayesian causal graph constructed by this engine uses activated semantic concepts as nodes in the graph. Each node represents an independent visual concept with human-interpretable semantics, such as "lung nodule," "vehicle / pedestrian," or "weld crack." Directed edges between nodes represent causal relationships, and their weights quantify the strength of these causal relationships. These causal edge weights are learned through a structural equation model.

[0034] Structural equation modeling (SEM) is a powerful statistical modeling framework that can handle multiple dependent and independent variables simultaneously and allows researchers to specify causal paths between variables. In this system, the SEM is trained on a historical decision dataset. This dataset contains a large number of original images, the semantic concepts activated by the system for these images, and the final decision outcomes given by domain experts. During training, the SEM uses maximum likelihood estimation to fit the response function of each concept node to its parent node. The basic principle of maximum likelihood estimation is to find a set of model parameters that maximize the probability of observed data occurrences. In this way, the model can learn the quantitative impact between various semantic concepts and their influence on the final decision outcome. For example, it can learn how the occurrence of "lung nodules" causally affects the decision of "cancer diagnosis," while "image noise" may only be correlated but not causally related.

[0035] During the inference phase, the causal inference engine utilizes the learned Bayesian causal graph to calculate the average causal effect of each activated concept on the final decision variable. The average causal effect is a metric that quantifies the average impact of changing the activation state of a specific concept on the final decision outcome, while keeping all other variables constant. This calculation typically involves intervention operations on the causal graph, i.e. The engine simulates virtual interventions in specific concepts to infer their causal impact on outcomes, rather than simply statistical associations. For example, by virtually "activating" or "deactivating" a concept in a graph, the engine can calculate how this intervention affects the probability distribution of the final decision. If the average causal effect value of an activated concept is greater than a preset causal significance threshold of 0.3, the concept is marked as a key causal concept. This threshold of 0.3 is based on domain knowledge and statistical analysis, aiming to ensure that only concepts that have a real and significant impact on the decision are identified as key causal factors. Physically, this means that the presence or absence of the concept can cause a change in the probability of the decision outcome by at least 30%.

[0036] The causal inference engine outputs a complete causal dependency graph, displaying activated semantic concept nodes, causal edges between them, and quantified edge weights. It also outputs a list of key causal concepts, explicitly identifying which concepts are the core drivers of the decision. Regarding anomaly handling, if the historical decision dataset contains mislabeled data or ambiguous causal definitions, potentially leading to poor training of the structural equation model, the engine will trigger a warning, suggesting data relabeling or expert review. If, during inference, the local structure of the causal graph becomes inconsistent or fails to converge, the engine will attempt to use heuristic algorithms for local correction or backtrack to the most recent stable graph state.

[0037] The counterfactual interpretation generation module, connected to the causal inference engine, aims to generate a counterfactual image variant with minimal perturbation, given the original image and the current decision outcome. Under this counterfactual image variant, the system outputs the opposite decision outcome. The key to this process is that the minimal perturbation is strictly limited to the spatial region corresponding to the key causal concepts identified by the causal inference engine, and is achieved through a gradient-guided pixel-level optimization algorithm.

[0038] The pixel-level optimization process performed in this module is as follows: First, using the original image as the initial value, the algorithm applies a differentiable perturbation to the spatial region corresponding to the key causal concepts identified by the causal inference engine. This means that the perturbation does not affect other parts of the image, but rather precisely targets the most decision-sensitive region. This spatial region information is provided by the semantic concept generation module, typically a specific bounding box or mask in the image. The perturbation is implemented through tiny additions and subtractions of pixel values, ensuring that the image content remains visually highly similar, with only subtle changes.

[0039] In order to guide the perturbation to effectively change the decision outcome and maintain the visual realism of the image, the system defines a loss function, which is a weighted sum of decision confidence loss and image fidelity loss.

[0040] The decision confidence loss incentivizes the model to output a high confidence level for the opposite category of the counterfactual image. Specifically, if the original decision is category A, the decision confidence loss penalizes cases where the counterfactual image is still classified as A and rewards cases where it is classified as a non-A (e.g., category B) with high confidence. This can be achieved using negative log-likelihood loss or cross-entropy loss, where the target category is set to the opposite of the original decision. Its mathematical form can be expressed as: ,in It is the opposite decision category. It is a counterfactual image variant.

[0041] Image fidelity loss employs the L2 norm constraint, limiting the perturbation amplitude to no more than 10% of the original pixel value. This loss term ensures that the generated counterfactual image is visually as close as possible to the original image, avoiding the generation of unnatural or excessively different "adversarial" images. The L2 norm is used to measure the original image... and counterfactual images The pixel-level differences between them. The constraint is... This can be simplified to ,in It is a perturbation applied to the image. It is a preset small constant representing the maximum permissible perturbation energy. This loss term ensures that the perturbation is visually imperceptible or difficult to detect.

[0042] The loss function combines these two objectives, balancing the speed of decision flipping with image fidelity requirements by adjusting their weights. The system iteratively updates the perturbation using the Adam optimizer. The Adam optimizer is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate of each parameter based on the first and second moment estimates of the gradient, thereby accelerating convergence and improving optimization stability. The optimization process continues until the decision result is flipped (i.e., the system's classification of the counterfactual image is opposite to the original image's classification) and the perturbation magnitude satisfies the L2 norm constraint. In each iteration, the optimizer fine-tunes the pixel values ​​within key causal concept regions based on the gradient information of the loss function, gradually pushing the image towards the other side of the decision boundary.

[0043] The output of the counterfactual explanation generation module is a counterfactual image variant that is highly similar to the original image but sufficient to alter the system's decision. This image provides intuitive and verifiable evidence for the decision comparison, helping human experts understand "how the decision would change if a small change were made to a key part of the image." Regarding anomaly handling, if the decision is too robust or the key causal concept region is too small, preventing a minimum perturbation from reversing the decision within a preset number of iterations, the module will abort the generation process and report that no valid counterfactual explanation has been found. Furthermore, if the generated counterfactual image exhibits significant visual distortion or artifacts, the system will initiate a post-processing algorithm to smooth or reconstruct it, ensuring its interpretability.

[0044] The human-machine collaborative verification interface connects the counterfactual explanation generation module and the semantic concept generation module. Its purpose is to simultaneously present the original images generated by the system, activated semantic concept regions, causal dependency graphs, and counterfactual image variants to domain experts in a visual format, and to receive feedback signals from experts regarding the reasonableness of the explanations. These feedback signals will be used to dynamically adjust the weight distribution of the visual concept space and the edge strength parameters of the causal graph, thereby achieving continuous optimization of the system's interpretability.

[0045] This interface provides a highly integrated visualization platform that presents a unified set of explanatory information to experts. The original image is typically displayed in the center of the interface. Activated semantic concept regions are overlaid on the original image as transparent masks or bounding boxes, while a list of corresponding semantic labels is displayed in the sidebar, sorted by confidence level, allowing experts to intuitively see the regions the model focuses on and their corresponding concepts. The causal dependency graph is presented graphically, with nodes representing activated semantic concepts, directed edges representing causal relationships, edge thickness or color depth quantifying causal strength, and arrows indicating the direction of causal action. Experts can click on nodes or edges through the interactive interface to view detailed information. Counterfactual image variants are usually displayed side-by-side with the original image or compared via a slider, highlighting the key perturbation regions that lead to decision reversals, as well as the different decision outcomes and their confidence levels corresponding to the two images, helping experts understand "what changes will change the decision."

[0046] This interface allows experts to label any causal edge in the causal graph with three categories: "accept," "reject," or "uncertain." For example, an expert can click on a causal edge and then select the corresponding feedback button.

[0047] "Acceptance" indicates that the expert believes the causal relationship is reasonable and consistent with domain knowledge.

[0048] "Rejection" indicates that the expert believes the causal relationship is unreasonable, erroneous, or violates common sense in the field.

[0049] "Uncertain" indicates that experts have reservations about the causal relationship and require more information or further verification.

[0050] These feedback signals, along with the contextual information at the time of feedback (e.g., original image ID, decision result, expert ID, timestamp), are accurately recorded and stored in an immutable log supported by distributed ledger technology, ensuring the authenticity and traceability of the feedback.

[0051] The system automatically triggers a parameter fine-tuning mechanism based on expert annotations. This mechanism uses an online learning method to update the internal parameters of the causal inference engine in real time, especially the edge weights of the causal graph.

[0052] If a causal edge is rejected multiple times, the system will reduce its edge weight in subsequent inferences to 50% of its initial value. Here, "multiple times" can be defined as a preset threshold, such as being rejected by more than 3 experts within a specific time window, or having a rejection rate exceeding 75%. This adjustment mechanism aims to weaken the influence of unreasonable causal relationships, so that they no longer act as strong causal factors in subsequent decisions.

[0053] Conversely, if a causal edge is "accepted" multiple times, the system will increase its edge weight to 150% of its initial value. This strengthens the causal relationship recognized by experts, allowing it to play a more important role in subsequent reasoning.

[0054] For causal edges marked as "uncertain", the system may maintain their weights or mark them as requiring more data or additional expert review, thereby triggering a data collection or expert consultation process.

[0055] The online learning approach for weight adjustment means that the system can optimize its internal parameters in real time and incrementally based on expert feedback without downtime or large-scale model retraining. This closed-loop feedback mechanism enables the system's interpretability to continuously evolve. By continuously absorbing valuable knowledge and experience from domain experts, the system can self-correct and self-improve, making its causal reasoning model increasingly consistent with human intuition and professional judgment, thereby greatly enhancing the credibility and practicality of the artificial intelligence system in high-risk application scenarios. Regarding anomaly handling, if significant internal inconsistencies in expert feedback are detected (e.g., different experts give contradictory and explicit feedback on the same edge), the system will trigger a conflict resolution mechanism, which may require higher-level expert arbitration or temporarily suspend the weight adjustment of that edge to avoid introducing erroneous learning signals.

[0056] This system further supports a dual-channel inference mode. In the normal mode, the system only outputs the final decision result, such as a classification label or predicted value. This mode is suitable for scenarios with high requirements for inference speed and efficiency, but relatively low requirements for interpretability, such as large-scale image batch processing or preliminary screening in real-time monitoring. In this mode, the system's computational resources are mainly focused on quickly generating accurate decision results, while omitting intermediate calculations and result generation related to interpretability, such as causal graph construction and counterfactual image generation.

[0057] In interpretable mode, the system synchronously outputs decision results, a list of activated semantic concepts, a causal dependency graph, and counterfactual image variants. This mode is suitable for high-risk, high-responsibility scenarios, such as medical diagnosis, industrial safety inspection, or autonomous driving decision assistance, where understanding the basis and logic behind the model's decisions is crucial. The activation of interpretable mode is dynamically triggered by an external safety policy controller based on the risk level of the application scenario. For example, in an autonomous driving system, when the vehicle is driving in complex road conditions or about to make a critical decision (such as emergency braking or lane changing), the safety policy controller will activate interpretable mode to provide decision-making support for the driver or remote operator; while in simple cruise mode, only the regular mode may be activated. The controller can determine mode switching based on preset rules, environmental perception data, or user commands, ensuring the system achieves an optimal balance between performance and transparency.

[0058] Furthermore, the system's visual concept space supports dynamic expansion. When the human-machine collaborative verification interface receives a new semantic concept definition and its example image submitted by an expert, the system automatically incorporates it into the concept ontology. For example, in an industrial quality inspection scenario, if a new defect pattern emerges, experts can upload an image example of the defect and name it (e.g., "new type of stress crack") via the interface. The system stores the new concept and its associated image data in the ontology and updates the knowledge graph. Subsequently, the system updates the mapping parameters of the semantic concept generation module through an incremental learning mechanism without retraining the entire system. Incremental learning allows the model to learn new concepts progressively without forgetting what it has already learned. This is typically achieved through mini-batch training on new data or by using techniques such as knowledge distillation, ensuring that the system maintains its interpretability and adaptability in the ever-evolving real world while significantly reducing maintenance and update costs.

[0059] This embodiment constructs a closed-loop, interpretable artificial intelligence analysis and reasoning system through the collaborative work of the aforementioned units. While maintaining high-performance image analysis capabilities, it significantly improves the transparency and traceability of decision-making logic. The system not only provides the final decision results but also clearly explains its reasoning process in the form of human-understandable semantic concepts, causal dependencies, and counterfactual comparisons. The human-machine collaborative verification interface realizes closed-loop feedback between expert knowledge and model reasoning, enabling the system's interpretability to continuously evolve and adapt to constantly changing external environments. This overall architecture provides solid technical support for the safe and reliable application of artificial intelligence in high-risk, high-responsibility fields, significantly improving the credibility and user acceptance of artificial intelligence systems.

[0060] Compared to existing technologies, this solution fundamentally overcomes the "black box" limitation prevalent in traditional deep learning models for image understanding tasks. Existing technologies often treat deep learning models as end-to-end mapping functions, with opaque internal decision-making processes, making them difficult to trust and deploy in critical application scenarios. This solution achieves white-box modeling of the decision-making process by constructing a complete interpretable path from raw pixels to human-understandable semantic concepts, and then to a causal logic chain. The combination of multi-scale feature extraction and a conceptual bottleneck network ensures that the system can both perceive image details and express them with structured semantics. The introduction of a causal inference engine enables the system to distinguish between statistical correlation and true causality, accurately identify key factors influencing decisions, and avoid being misled by irrelevant features or false associations. The counterfactual explanation generation mechanism provides intuitive and verifiable evidence for decision comparison, such as "if this small part of the image did not appear, the decision would be different," which greatly enhances human trust in AI judgments. The human-machine collaborative verification interface realizes a closed-loop feedback between expert knowledge and model reasoning, enabling the system's interpretability to continuously evolve and ensuring that the model's explanation always aligns with the domain expert's understanding. The overall architecture balances high-performance image analysis capabilities with highly transparent decision-making logic, providing solid technical support for the safe and reliable application of artificial intelligence in high-risk and high-responsibility fields, and has significant advantages and positive effects.

[0061] Example 2

[0062] This embodiment further expands and optimizes the internal mechanism of the causal inference engine based on Embodiment 1. The causal inference engine in Embodiment 1 primarily uses Bayesian causal graph models and structural equation models to model causal relationships and identify key concepts. However, in some complex scenarios, structural equation models are sensitive to linear relationships or pre-defined causal structures and may not be able to fully capture all potential nonlinear causal dependencies. Therefore, this embodiment introduces a reinforcement learning-based causal discovery and effect estimation method to enhance the robustness of the causal inference engine in handling nonlinear causal relationships and dynamic environments.

[0063] In this embodiment, the Bayesian causal graph constructed by the causal inference engine still uses activated semantic concepts as nodes, but the causal edge weights between nodes and the structure discovery process are obtained through a reinforcement learning-based search strategy. This reinforcement learning agent is trained to search for the optimal causal structure in the latent causal graph, thereby maximizing a certain reward function.

[0064] The reinforcement learning causal discovery process is as follows: First, the system defines an environment consisting of a set of activated semantic concepts and their observed data. An agent performs actions within this environment, each action corresponding to adding, deleting, or reversing a causal edge in the causal graph. The agent's goal is to learn a policy that minimizes the discrepancy between the fitted data and the discovered causal graph, while simultaneously penalizing the complexity of the causal graph.

[0065] The specific reward function can be designed as follows:

[0066]

[0067] in, It is the value of the reward function, and the agent's goal is to maximize this value; Given a causal graph The following observation data The log-likelihood of the causal graph measures the goodness of fit of the causal graph to the observed data. It is calculated using the Bayesian information criterion or the Akaike information criterion to ensure statistical validity. It is a preset hyperparameter used to balance data fitting and model complexity, and its value is determined based on experience or cross-validation. It is a cause-and-effect diagram The number of edges in the graph represents the complexity of the causal graph, and it is used to penalize overly complex models to avoid overfitting.

[0068] The reinforcement learning agent is trained on a historical decision dataset. The agent is a deep neural network, such as a policy gradient network, whose input is the current state representation of the causal graph and observation data, and whose output is a probability distribution for performing various graph operations (e.g., adding or deleting edges). Through interaction with the environment, the agent continuously adjusts its network parameters based on reward signals. When the agent is at a time step... Received observation data At that time, it will generate one action. For example, adding an edge from concept A to concept B. The environment then updates the causal graph based on this action and returns a reward of 1. , representing the quality of the new causal graph. The agent's strategy is to maximize the cumulative discount reward, i.e. ,in It is a discount factor, which controls the importance of future rewards.

[0069] During the inference phase, when new images are input and activated semantic concepts are generated, the causal inference engine runs a trained reinforcement learning agent to quickly converge to the optimal causal graph structure for the current data. Subsequently, the engine calculates the average causal effect value of each activated concept on the final decision variable. This calculation process can employ Pearl's do-calculus or G-computation frameworks, which, after causal discovery, can accurately estimate the intervention effect from the observed data, even with confounding variables. If the average causal effect value is greater than a preset causal significance threshold, the concept is marked as a key causal concept.

[0070] This embodiment overcomes the limitations of traditional structural equation modeling in handling nonlinear causal relationships and complex causal structures by introducing a reinforcement learning-based causal discovery mechanism. The reinforcement learning agent can autonomously explore and discover better causal structures within a vast graph space, enabling the causal inference engine to more accurately and comprehensively capture complex causal dependencies between semantic concepts in images. This enhanced causal modeling capability further improves the system's accuracy in identifying key causal factors, thus providing more reliable and in-depth interpretability for decision-making. In practical applications, such as medical diagnosis, this helps identify complex pathological causal chains that are difficult to discover using traditional statistical methods but are crucial to diagnostic results.

[0071] Example 3

[0072] This embodiment further optimizes and expands the functionality and implementation of the counterfactual explanation generation module based on Embodiments 1 and 2. Embodiment 1 primarily generates counterfactual image variants through pixel-level optimization, that is, by directly perturbing image pixels to change the model's decisions. However, in some scenarios, perturbing only at the pixel level may result in unnatural generated counterfactual images or semantic deviations from human-expected counterfactual explanations. Therefore, this embodiment introduces a counterfactual explanation generation method based on feature space perturbation, combined with generative adversarial networks (GANs) to generate higher-quality, more semantically coherent counterfactual explanations.

[0073] The pixel-level optimization process performed by the counterfactual interpretation generation module in this embodiment has been redesigned. Instead of directly applying perturbations to the pixel space of the original image, it performs differentiable perturbations in the feature space defined by the fused feature tensor output by the multi-scale feature extraction module.

[0074] The specific feature space optimization process is as follows: First, the system obtains the fused feature tensor of the original image from the multi-scale feature extraction module. Then, differentiable perturbations are applied to the feature dimensions or feature channels corresponding to the key causal concepts identified by the causal inference engine. This means that the perturbation no longer directly changes the pixels of the image, but rather alters the high-level semantic features of the image. This perturbation, applied to the feature space, better reflects the manipulation of the semantic content of the image, such as "reducing the visibility of a certain part in the image" or "enhancing a certain texture feature."

[0075] Similarly, the system defines a loss function, which is still a weighted sum of decision confidence loss and feature fidelity loss.

[0076] The objective of the decision confidence loss is the same as in Example 1, which is to induce the model to output a high confidence level for the opposite category of the counterfactual image reconstructed after feature space perturbation. Its mathematical form remains unchanged. ,in It is an image generator that converts from feature space to pixel space. It is the perturbed characteristic tensor.

[0077] The feature fidelity loss employs an L2 norm constraint, limiting the perturbation amplitude to no more than 5% of the original feature tensor value. Unlike pixel-level fidelity, this constraint operates on the feature space. This loss term ensures that the perturbed feature tensor maintains a high semantic similarity to the original feature tensor, avoiding the introduction of excessively large and unnatural changes into the feature space. Its mathematical form can be expressed as: ,in This is the original fused feature tensor. The L2 norm is chosen to ensure the smoothness and continuity of the feature perturbation.

[0078] After defining the loss function, the system iteratively updates the perturbation in the feature space using the Adam optimizer until the decision result is reversed and the perturbation magnitude satisfies the constraints.

[0079] Once a perturbation that satisfies the conditions is found in the feature space, the counterfactual interpretation generation module will utilize a pre-trained Generative Adversarial Network (GAN) as the image generator. This generator receives the perturbed feature tensor. As input, it is decoded and reconstructed into a counterfactual image variant in pixel space. Generative adversarial networks (GANs) can learn complex mappings from the feature space to the pixel distribution of real images. Therefore, the generated counterfactual images are often visually more natural, realistic, and semantically coherent, avoiding artifacts and unnaturalness that can result from direct pixel perturbations. During training, the generator is typically integrated with the encoder portion of a multi-scale feature extraction module to form an autoencoder structure, ensuring its efficient reconstruction of images from features.

[0080] This embodiment addresses the challenge of generating natural counterfactual explanations through direct pixel perturbation by combining feature space perturbation with generative adversarial networks (GANs) for image reconstruction. Perturbation in the feature space allows the system to manipulate image content at a higher semantic level, such as precisely "removing" or "modifying" a semantic concept in the image, rather than simply changing pixel colors. Combined with the image reconstruction capabilities of GANs, the generated counterfactual image variants not only successfully alter the model's decisions but also maintain a high degree of visual realism and semantic coherence, making these explanations easier for experts to understand and trust. For example, in autonomous driving scenarios, feature space perturbation can generate a counterfactual image that "misses a traffic sign," yet this image remains visually natural, aiding in understanding the model's reliance on traffic signs. This significantly improves the quality, understandability, and usability of counterfactual explanations.

Claims

1. An interpretable artificial intelligence analysis and reasoning system based on image computing, characterized in that, include: The image input module is used to receive raw image data and perform standardized preprocessing on it; The multi-scale feature extraction module, connected to the image input module, is used to extract local detail features, mesoscopic structural features and global semantic features of the image through parallel convolutional neural network branches, and fuse the three types of features in the channel dimension to form a unified multi-scale feature tensor. The semantic concept generation module, connected to the multi-scale feature extraction module, is used to map the multi-scale feature tensor to a set of predefined visual concept spaces with human-understandable semantics. The visual concept space is constructed from ontology knowledge bases in fields such as medical anatomical terms, industrial defect types, and traffic scene elements. Each visual concept corresponds to an interpretable semantic label and its spatial positioning region. The causal reasoning engine, connected to the semantic concept generation module, is used to model the causal dependencies between activated visual concepts based on the Bayesian causal graph model, identify key concept nodes that have a direct causal effect on the final decision result, and exclude interfering concepts that are only relevant but have no causal effect. The counterfactual interpretation generation module, connected to the causal inference engine, is used to generate a counterfactual image variant with minimal perturbation given the original image and the current decision result, so that the system outputs the opposite decision result under the variant. The minimal perturbation is limited to the spatial region corresponding to the key concept nodes identified by the causal inference engine and is implemented through a gradient-guided pixel-level optimization algorithm. The human-machine collaborative verification interface connects the counterfactual explanation generation module and the semantic concept generation module. It is used to simultaneously present the original image, the activated semantic concept region, the causal dependency graph, and the counterfactual image variants to the domain expert in a visual form, and to receive feedback signals from the expert on the rationality of the explanation. The feedback signals are used to dynamically adjust the weight distribution of the visual concept space and the edge strength parameters of the causal graph to achieve continuous interpretability optimization of the system.

2. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 1, characterized in that, The multi-scale feature extraction module comprises three parallel residual network branches, configured as shallow, medium, and deep networks, respectively. The shallow network contains two residual blocks for capturing high-frequency local features such as edges and textures; the medium network contains four residual blocks for extracting component-level structural features; and the deep network contains six residual blocks for encoding scene-level global semantics. The output feature maps of the three branches are upsampled or downsampled to a uniform spatial resolution, then concatenated along the channel dimension, and channel compression and feature fusion are performed through a 1×1 convolutional layer, outputting a fused feature tensor with a dimension of 128.

3. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 2, characterized in that, The semantic concept generation module adopts a concept bottleneck network architecture, which includes a fully connected layer and a concept activation function. The fully connected layer maps the fused feature tensor to an N-dimensional concept vector, where N is the total number of predefined visual concepts. The concept activation function is a piecewise linear threshold function. When a component in the concept vector exceeds a preset activation threshold of 0.7, the corresponding semantic concept is determined to be in an active state. The contribution region of the concept in the original image is located through backpropagation. The contribution region is generated by a gradient-weighted class activation mapping method with a spatial resolution of 1 / 4 of the input image.

4. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 3, characterized in that, The Bayesian causal graph constructed by the causal inference engine uses activated semantic concepts as nodes, and the causal edge weights between nodes are obtained through structural equation modeling. The structural equation model is trained based on historical decision datasets and uses maximum likelihood estimation to fit the response function of each concept node to its parent node. During the inference phase, the causal inference engine calculates the average causal effect value of each activated concept on the final decision variable. If the value is greater than the preset causal significance threshold of 0.3, the concept is marked as a key causal concept.

5. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 4, characterized in that, The pixel-level optimization process performed by the counterfactual explanation generation module is as follows: using the original image as the initial value, a differentiable perturbation is applied to the spatial region corresponding to the key causal concept; the loss function is defined as the weighted sum of decision confidence loss and image fidelity loss, where the decision confidence loss prompts the model to output a high confidence of the opposite category for the counterfactual image, and the image fidelity loss uses the L2 norm to constrain the perturbation amplitude to not exceed 10% of the original pixel value; the perturbation amount is iteratively updated through the Adam optimizer until the decision result is flipped and the perturbation amplitude meets the constraint condition.

6. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 5, characterized in that, The human-machine collaborative verification interface supports experts to label any causal edge in the causal graph with three categories: "accept", "reject", or "uncertain". The system automatically triggers a parameter fine-tuning mechanism based on the expert labels: if a causal edge is rejected multiple times, its edge weight in subsequent inference is reduced to 50% of the initial value; if it is accepted multiple times, its edge weight is increased to 150% of the initial value. The weight adjustment updates the internal parameters of the causal inference engine in real time through online learning.

7. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 6, characterized in that, The visual concept space supports dynamic expansion; when the human-machine collaborative verification interface receives a new semantic concept definition and its example image submitted by an expert, the system automatically incorporates it into the concept ontology library and updates the mapping parameters of the semantic concept generation module through an incremental learning mechanism, without having to retrain the entire system.

8. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 1, characterized in that, The system operates in a dual-channel inference mode; in normal mode, the system only outputs the final decision result. In interpretable mode, the system synchronously outputs decision results, a list of activated semantic concepts, a causal dependency graph, and a counterfactual image variant; the activation of interpretable mode is dynamically triggered by an external security policy controller based on the risk level of the application scenario.

9. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 4, characterized in that, The causal graph structure in the causal inference engine is obtained through a search strategy based on reinforcement learning; The reinforcement learning agent performs actions in an environment consisting of a set of activated semantic concepts and their observation data. Each action corresponds to adding, deleting, or reversing a causal edge in a causal graph. The agent learns the optimal causal structure by maximizing a reward function, which is determined by the goodness of fit of the causal graph to the observation data and the complexity of the causal graph.

10. The interpretable artificial intelligence analysis and reasoning system based on image computing according to claim 5, characterized in that, The counterfactual interpretation generation module performs a perturbation operation in the feature space; the system applies a differentiable perturbation to the feature dimension corresponding to the key causal concept in the fused feature tensor output by the multi-scale feature extraction module; the perturbed feature tensor is decoded into a counterfactual image variant in pixel space through a pre-trained adversarial generative network; the feature fidelity loss is constrained by the L2 norm to ensure that the perturbation amplitude does not exceed 5% of the original feature tensor value.

Images