Image calculation-based interpretable artificial intelligence analysis and reasoning system

Abstract

The invention belongs to the field of artificial intelligence, particularly relates to an interpretable artificial intelligence analysis and reasoning system based on image calculation, and aims to solve the problems that the AI image decision process is opaque and the causal logic is non-traceable. The system comprises a multi-scale feature extraction interface, a semantic concept generation interface, a causal inference engine interface, an anti-fact interpretation generation interface and a man-machine collaborative verification interface, and high-credibility image analysis is realized by constructing an interpretable path from pixels to semantic concepts and then to a causal 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 scheme of the present application is an explainable artificial intelligence analysis and reasoning system based on image calculation, comprising an image input module, a multi-scale feature extraction module, a semantic concept generation module, a causal reasoning engine, an counterfactual explanation generation module, and a man-machine collaborative verification interface; the image input module is used for receiving original image data and performing standardized preprocessing thereon; the multi-scale feature extraction module is connected to the image input module and is used for extracting local detail features, mesoscopic structure features and global semantic features of an image through parallel convolutional neural network branches respectively, fusing the three types of features in the channel dimension to form a unified multi-scale feature tensor; the semantic concept generation module is connected to the multi-scale feature extraction module and is used for mapping the multi-scale feature tensor to a group of predefined visual concept spaces with human understandable semantics, the visual concept spaces being constructed from ontology knowledge bases in the fields of medical anatomy terms, industrial defect types, traffic scene elements, etc., each visual concept corresponding to an explainable semantic label and its spatial positioning area; the causal reasoning engine is connected to the semantic concept generation module and is used for modeling the causal dependence relationship between activated visual concepts based on a Bayesian causal graph model, identifying key concept nodes having a direct causal effect on the final decision result, and excluding interference concepts having only correlation but no causal effect; the counterfactual explanation generation module is connected to the causal reasoning engine and is used for generating a counterfactual image variant under the minimum disturbance given the original image and the current decision result, so that the system outputs an opposite decision result under the variant, the minimum disturbance being limited to the spatial area corresponding to the key concept nodes identified by the causal reasoning engine and being realized through a gradient-guided pixel-level optimization algorithm; the man-machine collaborative verification interface is connected to the counterfactual explanation generation module and the semantic concept generation module and is used for synchronously presenting the original image, the activated semantic concept area, the causal dependence graph, and the counterfactual image variant to a field expert in a visualized form, and receiving feedback signals of the expert on the rationality of the explanation, the feedback signals being used for dynamically adjusting the weight distribution of the visual concept space and the edge strength parameters of the causal graph, realizing continuous explainability optimization of the system.

[0007] Further, the multi-scale feature extraction module comprises three parallel residual network branches configured as a shallow network, a middle network and a deep network respectively; the shallow network comprises 2 residual blocks for capturing high-frequency local features such as edges and textures; the middle network comprises 4 residual blocks for extracting part-level structure features; the deep network comprises 6 residual blocks for encoding scene-level global semantics; the output feature maps of the three branches are upsampled or downsampled to a unified spatial resolution, spliced in the channel dimension, and compressed in the channel and fused through a 1x1 convolution layer, outputting a fused feature tensor with a dimension of 128.

[0008] Further, 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, N being 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 activated state, and the contribution area of the concept in the original image is located through back propagation, and the contribution area is generated through a gradient-weighted class activation mapping method, and the spatial resolution is 1 / 4 of the input image.

[0009] Further, the causal reasoning engine constructs a Bayesian causal graph with activated semantic concepts as nodes, and the causal edge weights between nodes are obtained through a structural equation model; the structural equation model is trained based on a historical decision dataset, and a maximum likelihood estimation method is used to fit the response function of each concept node to its parent node; in the reasoning phase, the causal reasoning engine calculates the average causal effect value of each activated concept on the final decision variable, and if the value is greater than a preset causal significance threshold of 0.3, the concept is marked as a key causal concept.

[0010] Further, the counterfactual explanation generation module performs a pixel-level optimization process as follows: taking the original image as the initial value, applying a differentiable perturbation in the spatial area corresponding to the key causal concept; defining a loss function as the weighted sum of decision confidence loss and image fidelity loss, wherein the decision confidence loss encourages the model to output a high confidence of the opposite class for the counterfactual image, and the image fidelity loss uses an L2 norm to constrain the perturbation amplitude to be no more than 10% of the original pixel value; the perturbation amount is iteratively updated by an Adam optimizer until the decision result is reversed and the perturbation amplitude meets the constraint condition.

[0011] Further, the human-machine collaborative verification interface supports experts to make three types of annotations, i.e., "accept", "reject" or "uncertain", on any causal edge in the causal graph; the system automatically triggers a parameter fine-tuning mechanism according to the expert annotations: if a causal edge is rejected multiple times, its edge weight in subsequent reasoning 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 reasoning engine in real time through online learning.

[0012] Further, 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 includes it in the concept ontology library and updates the mapping parameters of the semantic concept generation module through an incremental learning mechanism, without the need to retrain the entire system.

[0013] Further, the system runs in a dual-channel inference mode; in a regular mode, the system only outputs the final decision result; in an explainable mode, the system synchronously outputs the decision result, the activated semantic concept list, the causal dependence graph, and the counterfactual image variant; the activation of the explainable mode is dynamically triggered by an 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 application are that: The present application fundamentally breaks through the black box limitation of traditional deep learning models by constructing a complete explainable path from raw pixels to human understandable semantic concepts and then to causal logic chains; the combination of multi-scale feature extraction and concept bottleneck network ensures that the system can perceive image details and express them in structured semantics; the introduction of the causal reasoning engine enables the system to distinguish between correlation and causality, accurately identify key factors affecting decisions, and avoid being misled by irrelevant features; the counterfactual explanation generation mechanism provides intuitive and verifiable decision comparison evidence, greatly improving human trust in AI judgment; the human-machine collaborative verification interface realizes closed-loop feedback of expert knowledge and model reasoning, making the explainability of the system have the ability to continuously evolve; the overall architecture takes into account both high-performance image analysis capabilities and high-transparency decision logic, providing a solid technical support for the safe and reliable application of artificial intelligence in high-risk and high-responsibility fields. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a schematic diagram of the overall technical scheme of the present application; Figure 2 is a schematic diagram of the core principle framework of the cooperation between the causal reasoning engine and the counterfactual explanation generation module in the present application. DETAILED DESCRIPTION

[0016] Embodiment 1 Please refer to Figure 1 and Figure 2 The present application provides an explainable artificial intelligence analysis and reasoning system based on image computing, the core purpose of which is to solve the "black box" problem commonly existing in existing image understanding artificial intelligence systems, i.e., the lack of transparency in model decision-making process, the untraceability of causal logic, and the difficulty for humans to understand the reasoning basis, etc. technical challenges. By constructing a complete explainable path from raw image pixel data to human understandable semantic concepts and then to causal logic chains, the present system fundamentally breaks through the limitations of traditional deep learning models in image understanding tasks.

[0017] The system architecture includes an image input module, a multi-scale feature extraction module, a semantic concept generation module, a causal reasoning engine, an counterfactual explanation generation module, and a human-machine collaborative verification interface. The overall technical process of the system starts with the image input module receiving and preprocessing the original image data, and then the multi-scale feature extraction module extracts multi-dimensional features from the image. The extracted features are sent to the semantic concept generation module, which maps them to visual concepts with human-understandable semantics. These activated visual concepts and their spatial positioning information are then used by the causal reasoning engine to build causal dependency relationships, identifying key concept nodes that have a direct causal effect on the final decision result. 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 presents the original image, semantic concepts, causal dependency graph, and counterfactual image variants to domain experts simultaneously, and receives feedback signals to continuously optimize the explainability of the system. The system also supports a dual-channel reasoning mode, which can dynamically switch between regular and explainable modes according to the risk level of the application scenario, and the visual concept space has dynamic expansion capability.

[0018] Image input module, aims to receive raw image data from various sources and perform necessary standardization preprocessing to ensure that subsequent processing units can utilize the data stably and efficiently. This module is an integrated data access and preprocessing platform that supports the input of multiple image data formats, such as JPEG, PNG, DICOM medical image format, and various sensor-collected raw pixel streams. When the module receives raw image data, a series of automated preprocessing procedures will be initiated. The first procedure is image format analysis and decoding, which converts raw binary data into an operable pixel matrix. Subsequently, data integrity checks are performed, such as checking whether the file header, image size meets expectations, and whether there are data damage or loss; if abnormalities are detected, the system will generate an error report and attempt data recovery or perform abnormal 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 particularly important for sensor-collected images or low-quality images. Subsequently, brightness and contrast adjustment is performed, such as using histogram equalization techniques to enhance the visual distinguishability and feature differentiation of the image. Size normalization is a key step, the module uniformly scales all images to a pre-set standard spatial resolution, such as 256 pixels by 256 pixels or 512 pixels by 512 pixels, commonly used interpolation algorithms include bilinear interpolation or bicubic interpolation, to ensure the input size consistency of the feature extraction module. Pixel value normalization is to map the pixel intensity value to a standard range, such as 0 to 1, or to perform zero-mean unit variance standardization, which helps stabilize the neural network training process and improve convergence speed. For specific application scenarios, 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 performed according to the predefined processing pipeline to ensure that the output image data is a uniform format, standardized and reliable quality tensor, ready for input into the multi-scale feature extraction module. In terms of exception handling, if the image is severely damaged by external factors and cannot be repaired, the module will trigger a rejection processing mechanism and return an error code and log information to the upper-level caller, preventing the pollution of subsequent processing procedures.

[0019] Multi-scale feature extraction module, connected to the image input module, aims to capture local detailed features, medium structural features and global semantic features of the image from different abstraction levels through parallel convolutional neural network branches, and efficiently fuse these heterogeneous features in the channel dimension to form a unified and information-rich multi-scale feature tensor. The design philosophy of this module is to simulate the cognitive process of the human eye when observing objects, which can focus on both fine details and overall contours. The module contains three parallel residual network branches, configured as shallow network, medium network and deep network respectively, each branch focuses on extracting features of a specific scale.

[0020] The shallow network contains 2 residual blocks, whose main task is to capture high-frequency local features such as edges, textures, etc. in the image. The internal structure of each residual block usually includes two convolutional layers, two batch normalization layers, and two activation function layers, and adopts a skip connection mechanism to directly add the input to the output of the block, in order to alleviate the gradient vanishing problem in deep network training and improve the feature representation ability. The size of the convolution kernel of the shallow network is usually small, such as 3 by 3, aiming to accurately capture the pixel gradient changes in the local area of the image, such as detecting the outline, texture details of the object. The network outputs feature maps with high spatial resolution and fewer channels, which directly reflect the original visual elements of the image. When the data flows through the shallow network, each layer of convolutional operation extracts edge information in different directions and frequencies, the batch normalization layer stabilizes the distribution of these features, and the activation function such as the rectified linear unit introduces a nonlinear transformation, which can be combined into more complex local patterns.

[0021] The middle network contains 4 residual blocks, mainly used to extract component-level structural features. Compared with the shallow network, the middle network has a larger receptive field, which can aggregate information from multiple local features to form an abstract representation of the object components in the image. For example, in a face image, the middle network may identify features such as eyes, nose, and mouth; in an industrial quality inspection scenario, it may identify the shape of a screw and the direction of a weld. By increasing the number of residual blocks, the middle network can perform deeper nonlinear transformations and feature abstraction. The internal down-sampling operation, such as through a convolutional layer with a step of 2, gradually reduces the spatial resolution of the feature map while increasing the number of channels, thereby encoding information at a higher semantic level.

[0022] The deep network contains 6 residual blocks, whose core function is to encode scene-level global semantics. The receptive field of the deep network is the largest, which can understand the overall content, background information, and spatial relationship between different components of the image. For example, in an autonomous driving scenario, the deep network can identify "highway scenes" and "urban road scenes"; in medical images, it can identify the overall distribution of "tumor regions" and "healthy tissue regions". The deep network further stacks residual blocks and performs down-sampling to highly compress the image information into feature maps with low spatial resolution and high channel number, which contain the highest level of abstract semantic information of the image.

[0023] The output feature maps of the three parallel branches can have different spatial resolutions, so they need to be upsampled or downsampled to a predetermined standard spatial resolution, for example, using bilinear interpolation or a deconvolutional layer. After the resolution is unified, the feature maps are concatenated in the channel dimension. The concatenated feature tensor contains rich information from different scales and different levels of abstraction. Finally, the concatenated tensor is passed through a 1x1 convolutional layer for channel compression and feature fusion. The role of this 1x1 convolutional layer is to reduce the dimension and further mix information of different scales, while introducing a nonlinear transformation to enhance the feature representation capability, and finally output a fusion feature tensor with a dimension of 128. This fusion feature tensor is a high-dimensional vector that integrates local details, mesoscopic structures, and global semantic information of the image, providing comprehensive input for subsequent semantic concept generation. In terms of parameter management, the weights of each residual network branch can be pre-trained on a large-scale image dataset, such as ImageNet, to obtain good general feature extraction capability, and then fine-tuned on a specific task dataset to adapt to the specific application requirements of the system. Weight updates use standard gradient descent optimization algorithms, combined with batch normalization and weight decay techniques to ensure the stability of the training and the generalization ability of the model.

[0024] The semantic concept generation module is connected to the multi-scale feature extraction module, and its purpose is to map the fusion feature tensor output by the multi-scale feature extraction module to a set of predefined visual concept spaces with human-understandable semantics, and generate corresponding semantic labels and spatial positioning areas in the original image for each activated visual concept. The core of this module is to convert the abstract features learned by the neural network into high-level semantic concepts that humans can intuitively understand, thereby achieving model interpretability.

[0025] This module uses a concept bottleneck network architecture. This architecture first includes a fully connected layer, which takes the fusion feature tensor with a dimension of 128 as input. Before inputting the fully connected layer, the fusion feature tensor usually undergoes a global average pooling operation to compress its spatial dimension to a single vector, effectively passing the global information of 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 visual concepts predefined by the system. Each component corresponds to a specific visual concept, and its value represents the intensity or confidence of the concept being activated.

[0026] After generating the N-dimensional concept vector, the system applies a concept activation function. The activation function takes the form of a piecewise linear threshold function. Specifically, when a component of the concept vector exceeds a pre-set activation threshold of 0.7, the corresponding semantic concept is determined to be in an activated state. The threshold setting aims to filter out low-confidence concepts and only keep those semantic concepts with strong evidence support. For example, if a component of the concept vector represents “lung nodule” and its value reaches 0.8, the semantic concept of “lung nodule” is considered to be activated. 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 implements the binary activation of concepts while preserving the gradient.

[0027] Once a semantic concept is determined to be in an activated state, the system locates the contribution region of the concept in the original image through the backpropagation mechanism. Specifically, the gradient-weighted class activation mapping method, Grad-CAM, is adopted. Grad-CAM calculates the gradient of the target concept on the last layer of convolutional feature maps, and performs global average pooling on these gradients to generate a weight vector. Then, the weight vector is weighted and summed with the last layer of convolutional feature maps, and processed through an activation function to generate a heat map. This heat map clearly indicates the region in the original image that contributes most to the activation of a specific concept. The spatial positioning region generated by Grad-CAM has a resolution of 1 / 4 of the input image, which means it can provide relatively fine regional positioning information, but not pixel-level. For example, for an input image of 256 pixels by 256 pixels, the resolution of the positioning region is 64 pixels by 64 pixels. This region can be a rectangular bounding box or a pixel-level mask, depending on the accuracy requirements of subsequent analysis.

[0028] The visual concept space is constructed from a pre-defined domain ontology knowledge base. This knowledge base is a structured knowledge system that contains all relevant concepts, hierarchical relationships between concepts, and attributes within a specific application domain. For example, in the medical diagnosis domain, the knowledge base may contain "medical anatomical terms" (such as "lung", "heart", "liver"), "disease types" (such as "pneumonia", "tumor"), and "imaging features" (such as "ground glass opacity", "cavity"). In the industrial quality inspection domain, it may contain "industrial defect types" (such as "scratch", "crack", "foreign matter"), and "component names" (such as "bolt", "weld"). In the traffic scene element domain, it may contain "vehicle types", "pedestrians", "traffic signs", and "lane lines". These concepts and their semantic labels are defined and labeled by domain experts, and managed and maintained by ontology tools to ensure their understandability and normativity in human cognition. Each visual concept not only has a semantic label, but also may be associated with metadata describing its characteristics, such as typical morphology, size range, etc. When new information is collected, the system passes the corresponding semantic label and spatial positioning area to the causal reasoning engine as the basis for modeling causal relationships.

[0029] The causal reasoning engine is connected to the semantic concept generation module, and its purpose is to model the causal dependence relationships between the visual concepts activated by the semantic concept generation module based on the 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 result, and exclude interference concepts that only have relevance but no causal effect, thereby providing more accurate and reliable decision explanations.

[0030] The Bayesian causal graph constructed by this engine takes the 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", and "weld crack". The directed edges between nodes represent causal relationships, and their weights quantify the strength of these causal relationships. These causal edge weights are obtained through structural equation model learning.

[0031] Structural equation modeling is a powerful statistical modeling framework that can handle multiple dependent and independent variables simultaneously and allows researchers to specify causal pathways between variables. In this system, the structural equation model is trained on a historical decision dataset. This dataset contains a large number of raw images, semantic concepts activated by the system for these images, and the final decision results given by domain experts. During training, the structural equation model uses maximum likelihood estimation to fit the response function of each concept node to its parent nodes. The basic principle of maximum likelihood estimation is to find a set of model parameters that maximize the probability of observing the data under these parameters. In this way, the model can learn the quantitative influence between each semantic concept and its impact on the final decision result. For example, it can learn how the presence of "lung nodules" causally affects the decision of "cancer diagnosis," while "image noise" may only be related but not causally related.

[0032] In the reasoning phase, the causal reasoning engine uses the learned Bayesian causal graph to calculate the average causal effect value of each activated concept on the final decision variable. The average causal effect value is a measure that quantifies the average impact on the final decision result if the activation state of a particular concept is changed while all other variables remain unchanged. This calculation process usually involves intervention operations on the causal graph, i.e. , which simulates a virtual intervention on a specific concept to infer its causal impact on the result rather than just a statistical association. For example, by virtually "activating" or "deactivating" a concept in the 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 the preset causal significance threshold of 0.3, the concept is marked as a key causal concept. This threshold of 0.3 is set based on domain knowledge and statistical analysis, aiming to ensure that only those concepts that have a real and significant impact on the decision are identified as key causal factors. Its physical meaning is that the presence or absence of the concept can cause at least a 30% change in the probability of the decision result.

[0033] The output of the causal reasoning engine includes a complete causal dependency graph showing activated semantic concept nodes, causal edges between them, and quantified edge weights; at the same time, it also outputs a list of key causal concepts, clearly indicating which concepts are the core driving factors of this decision. In terms of exception handling, if the historical decision dataset has label errors or ambiguous causal relationship definitions, which may lead to poor training of the structural equation model, the engine will trigger a warning and prompt the need for re-labeling or expert review of the data. If inconsistencies or non-convergence occur in the local structure of the causal graph during reasoning, the engine will attempt to use heuristic algorithms for local correction or backtrack to the last stable graph state.

[0034] The counterfactual explanation generation module is connected to the causal reasoning engine, aiming to generate a counterfactual image variant under the premise of the given original image and the current decision result. Under this counterfactual image variant, the system outputs an opposite decision result. The key to this process is that the minimum perturbation is strictly limited to the spatial regions corresponding to the key causal concepts identified by the causal reasoning engine, and is realized through a gradient-guided pixel-level optimization algorithm.

[0035] The pixel-level optimization process performed by this module is as follows: first, taking the original image as the initial value, the algorithm will apply a differentiable perturbation within the spatial regions corresponding to the key causal concepts identified by the causal reasoning engine. This means that the perturbation will not affect other parts of the image, but will precisely act on the most sensitive areas of the decision. This spatial region information is provided by the semantic concept generation module, usually a specific bounding box or mask in the image. The perturbation is realized in the form of a small addition or subtraction of pixel values, ensuring that the image content still maintains a high degree of similarity in vision, only with subtle changes.

[0036] In order to guide the perturbation to effectively change the decision result and maintain the visual fidelity of the image, the system defines a loss function composed of the weighted sum of the decision confidence loss and the image fidelity loss.

[0037] The decision confidence loss encourages the model to output a high confidence of the opposite class for the counterfactual image. Specifically, if the original decision is class A, the decision confidence loss will penalize the case where the counterfactual image is still classified as A, and reward the case where it is classified as non-A (e.g., class B) with high confidence. This can be achieved through a negative log-likelihood loss or a cross-entropy loss, where the target class is set to the opposite class of the original decision. Its mathematical form can be expressed as , where is the opposite decision class, is the counterfactual image variant.

[0038] The image fidelity loss uses the L2 norm to constrain the perturbation amplitude to no more than 10% of the original pixel value. This loss term ensures that the generated counterfactual image is as close as possible to the original image in vision, avoiding the generation of "adversarial" images that are unnatural or too different from the original image. The L2 norm is used to measure the pixel-level difference between the original image and the counterfactual image . The constraint condition is , which can be simplified as , where is the perturbation applied to the image, is a small constant preset to represent the maximum allowed perturbation energy. This loss term ensures that the perturbation is visually imperceptible or difficult to perceive.

[0039] 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.

[0040] 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.

[0041] 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.

[0042] The interface provides a highly integrated visualization platform that unifies various interpretation information and presents them to the experts. The original image is usually displayed in the center of the interface. The activated semantic concept regions are overlaid on the original image in the form of transparent masks or bounding boxes, while the corresponding semantic label list is displayed in the sidebar, sorted by confidence, allowing the expert to intuitively see the areas of interest of the model and their corresponding concepts. The causal dependency graph is presented in a graphical manner, with nodes representing activated semantic concepts and directed edges representing causal relationships. The thickness or color depth of the edges quantifies the causal strength, and the arrow direction indicates the causal direction. Experts can click on nodes or edges through the interactive interface to view detailed information. Counterfactual image variants are usually displayed side byside with the original image or compared through a sliding bar, while highlighting the key perturbation areas that lead to decision reversal, as well as the different decision results and their confidence levels for the two images, helping experts understand "what changes will change the decision."

[0043] The interface supports experts to label any causal edge in the causal graph as "accept", "reject" or "uncertain". For example, the expert can click on a causal edge and then select the corresponding feedback button.

[0044] "Accept" means that the expert believes that the causal relationship is reasonable and consistent with domain knowledge.

[0045] "Reject" means that the expert believes that the causal relationship is unreasonable, wrong or violates common sense in the field.

[0046] "Uncertain" means that the expert has reservations about the causal relationship and needs more information or further verification.

[0047] These feedback signals and the context information at the time of feedback (e.g. original image ID, decision result, expert ID, timestamp) will be accurately recorded and stored in a tamper-proof log supported by distributed ledger technology, ensuring the authenticity and traceability of the feedback.

[0048] The system automatically triggers a parameter fine-tuning mechanism based on expert annotations, which updates the internal parameters of the causal reasoning engine in real time, especially the edge weights of the causal graph, using online learning.

[0049] If a causal edge is rejected multiple times, the system will reduce its edge weight in subsequent reasoning to 50% of the initial value. Here, "multiple times" can be defined as a preset threshold, such as being rejected by more than 3 experts within a certain time window, or a rejection rate exceeding 75%. This adjustment mechanism aims to weaken the impact of unreasonable causal relationships, so that they are no longer considered as strong causal factors in subsequent decisions.

[0050] Conversely, if a causal edge is accepted multiple times, the system will boost its edge weight to 150% of the initial value. This reinforces the causal relationship that is endorsed by experts, making it play a more important role in subsequent reasoning.

[0051] For causal edges marked as “uncertain”, the system can either maintain their weights unchanged, or mark them as requiring more data or additional expert review, triggering data collection or expert consultation processes.

[0052] The online learning approach of weight adjustment means that the system can optimize its internal parameters in real time and incrementally according to expert feedback without the need for downtime or large-scale model retraining. This closed-loop feedback mechanism enables the explainability of the system to evolve continuously. By continuously absorbing the valuable knowledge and experience of domain experts, the system can self-correct and self-improve, making its causal reasoning model more and more consistent with human intuition and professional judgment, thereby greatly enhancing the credibility and practicality of artificial intelligence systems in high-risk application scenarios. In terms of exception handling, if significant internal inconsistencies are detected in expert feedback (for example, different experts give opposite explicit feedback on the same edge), the system will trigger a conflict resolution mechanism, which may require higher-level experts to arbitrate or temporarily suspend the weight adjustment of the edge to avoid introducing incorrect learning signals.

[0053] The system further supports a dual-channel reasoning mode. In the regular mode, the system only outputs the final decision result, such as a classification label or a prediction value. This mode is suitable for scenarios where reasoning speed and efficiency are highly valued, while explainability is relatively low, such as large-scale image batch processing or preliminary screening in real-time monitoring. In this mode, the system's computing resources are mainly concentrated on quickly generating accurate decision results, while omitting intermediate calculations and result generation related to explainability, such as causal graph construction and counterfactual image generation.

[0054] In the explainable mode, the system simultaneously outputs the decision result, the list of activated semantic concepts, the causal dependency graph, and the 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 of the model's decision is crucial. The activation of the explainable 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 on complex roads or is about to make a critical decision (such as emergency braking or lane changing), the safety policy controller will activate the explainable mode to provide decision-making basis for the driver or remote operator; while in simple cruise mode, only the regular mode may be enabled. The controller can decide mode switching based on pre-set rules, environmental perception data, or user instructions, ensuring the best balance between performance and transparency.

[0055] Furthermore, the visual concept space of the present system supports dynamic expansion. When the human-machine collaborative verification interface receives a new semantic concept definition and its example images submitted by an expert, the system automatically incorporates it into the concept ontology library. For example, in the industrial quality inspection scenario, if a new type of defect pattern appears, the expert can upload the image example of the defect and name it (e.g., "new stress crack") through the interface. The system will store the new concept and its associated image data in the ontology library and update 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 gradually learn new concepts without forgetting previously learned knowledge. This is typically achieved through small-batch training on new data or using techniques such as knowledge distillation, ensuring that the system maintains its explainability and adaptability in a constantly evolving real world, while significantly reducing the cost of maintenance and updates.

[0056] The present embodiment, through the cooperative work of the above-mentioned units, constructs a closed-loop explainable artificial intelligence analysis and reasoning system. It not only maintains high-performance image analysis capabilities but also significantly improves the transparency and traceability of decision logic. The system can not only provide the final decision result but also clearly explain its reasoning process in the form of human-understandable semantic concepts, causal dependency relationships, and counterfactual comparisons. The human-machine collaborative verification interface realizes the closed-loop feedback of expert knowledge and model reasoning, enabling the explainability of the system to continuously evolve and adapt to changing external environments. This overall architecture provides solid technical support for the safe and reliable application of artificial intelligence in high-risk and high-responsibility fields, significantly improving the credibility and user acceptance of artificial intelligence systems.

[0057] Compared with the prior art, the scheme fundamentally breaks through the "black box" limitation of traditional deep learning models in image understanding tasks. The prior art often regards the deep learning model as an end-to-end mapping function, and the internal decision-making process is not transparent, which makes it difficult to be trusted and deployed in key application scenarios. The scheme realizes the white-box of the model decision-making process by building a complete interpretable path from the original pixels to the human understandable semantic concepts and then to the causal logic chain. The combination of multi-scale feature extraction and concept bottleneck network ensures that the system can perceive image details and express structured semantics. The introduction of the causal reasoning engine enables the system to distinguish between statistical correlation and real causality, accurately identify key factors affecting decisions, and avoid being misled by irrelevant features or false correlations. The counterfactual explanation generation mechanism provides intuitive and verifiable decision comparison evidence, such as "if this small part of the image does not appear, the decision will be different", which greatly improves the trust of humans in artificial intelligence judgment. The man-machine cooperative verification interface realizes the closed-loop feedback of expert knowledge and model reasoning, making the system's explainability have the ability to evolve continuously, ensuring that the model explanation always keeps consistent with the cognition of domain experts. The overall architecture balances high-performance image analysis capability and high-transparency decision 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.

[0058] Embodiment 2 Based on Embodiment 1, the internal mechanism of the causal reasoning engine is further expanded and optimized in this embodiment. The causal reasoning engine in Embodiment 1 mainly models the causal relationship and identifies key concepts based on the Bayesian causal graph model and the structural equation model. However, in some complex scenarios, the structural equation model is more sensitive to linear relationships or pre-set causal structures, and may not be able to fully capture all potential nonlinear causal dependencies. Therefore, this embodiment introduces a causal discovery and effect estimation method based on reinforcement learning to enhance the robustness of the causal reasoning engine in handling nonlinear causal relationships and dynamic environments.

[0059] The Bayesian causal graph constructed by the causal reasoning engine in this embodiment still takes activated semantic concepts as nodes, but the causal edge weights and structure discovery process between nodes are obtained through a reinforcement learning-based search strategy. The reinforcement learning agent is trained to search for the optimal causal structure in the potential causal graph, thereby maximizing a certain reward function.

[0060] The reinforcement learning causal discovery process is as follows: First, the system defines an environment consisting of the set of activated semantic concepts and their observation data. The agent performs actions in the environment, each corresponding to adding, removing, or reversing a causal edge in the causal graph. The agent's goal is to learn a policy that minimizes the difference between the fitted data and the discovered causal graph while penalizing the complexity of the causal graph.

[0061] The specific reward function can be designed as: where, is the value of the reward function, which the agent aims to maximize; is the log-likelihood of the observation data given the causal graph , which measures the goodness-of-fit of the causal graph to the observation data, computed using the Bayesian Information Criterion or Akaike Information Criterion to ensure statistical validity; is a pre-set hyperparameter that balances data fitting and model complexity, whose value is determined empirically or through cross-validation; is the number of edges in the causal graph , representing the complexity of the causal graph, which penalizes overly complex models to avoid overfitting.

[0062] The training of the reinforcement learning agent is based on a historical decision dataset. The agent is a deep neural network, such as a policy gradient network, whose input is the state representation of the current causal graph and the observation data, and whose output is the probability distribution of performing various graph operations (e.g., adding edges, removing edges). Through interaction with the environment, the agent continuously adjusts its network parameters according to the reward signal. When the agent receives observation data at a time step , it generates an action , such as adding an edge from concept A to concept B. The environment then updates the causal graph according to this action and returns a reward , indicating the quality of the new causal graph. The agent's policy is to maximize the cumulative discounted reward, i.e. where is the discount factor, controlling the importance of future rewards.

[0063] ​In the inference phase, when new images are input and activated semantic concepts are generated, the causal reasoning engine runs the trained reinforcement learning agent to quickly converge to the optimal causal graph structure under 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 adopt Pearl's do-calculus or G-computation framework, which can accurately estimate the intervention effect from observed data even in the presence of confounding variables after causal discovery. If the average causal effect value is greater than the preset causal significance threshold, the concept is marked as a key causal concept.

[0064] The embodiment introduces a reinforcement learning-based causal discovery mechanism to overcome the limitations of traditional structural equation models in handling nonlinear causal relationships and complex causal structures. The reinforcement learning agent can autonomously explore and discover better causal structures in a vast graph space, enabling the causal reasoning engine to more accurately and comprehensively capture the complex causal dependency relationships between semantic concepts in images. This enhanced causal modeling capability further improves the accuracy of the system in identifying key causal factors, thereby providing more reliable and in-depth explainability 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 for diagnosis results.

[0065] Embodiment 3 The embodiment further optimizes and extends the functions and implementation of the counterfactual explanation generation module based on embodiments 1 and 2. In embodiment 1, counterfactual image variants are mainly generated through pixel-level optimization, i.e., by directly perturbing image pixels to change the model's decision. However, in some scenarios, perturbation only at the pixel level may result in counterfactual images that are not natural enough or deviate from the semantic expectations of human counterfactual explanations. To this end, the embodiment introduces a counterfactual explanation generation method based on feature space perturbation and combines adversarial generative network technology to generate higher-quality and more semantically coherent counterfactual explanations.

[0066] The pixel-level optimization process performed by the counterfactual explanation generation module in this embodiment is redesigned. Instead of directly applying perturbation in the pixel space of the original image, it performs differentiable perturbation in the feature space defined by the fusion feature tensor output by the multi-scale feature extraction module.

[0067] 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. Subsequently, a differentiable perturbation is applied on the feature dimension or feature channel corresponding to the key causal concept identified by the causal reasoning engine. This means that the perturbation is no longer directly changing the pixels of the image, but changing the high-level semantic features of the image. This perturbation acts on the feature space, and can better reflect the manipulation of the semantic content of the image, such as "reducing the visibility of a certain component in the image" or "enhancing a certain texture feature".

[0068] Similarly, the system defines a loss function, which is still a weighted sum of the decision confidence loss and the feature fidelity loss.

[0069] The goal of the decision confidence loss is the same as in Embodiment 1, that is, to encourage the model to output a high confidence of the opposite class for the counterfactual image reconstructed after the feature space perturbation. Its mathematical form remains , where is an image generator from the feature space to the pixel space, is the perturbed feature tensor.

[0070] The feature fidelity loss uses the L2 norm to constrain the perturbation amplitude to not exceed 5% of the original feature tensor value. Unlike pixel-level fidelity, the constraint here acts on the feature space. This loss term ensures that the perturbed feature tensor and the original feature tensor remain highly similar in semantics, avoiding the introduction of too large and unnatural changes in the feature space. Its mathematical form can be expressed as , where is the original fused feature tensor. The L2 norm is chosen to ensure the smoothness and continuity of the feature perturbation.

[0071] After defining the loss function, the system iteratively updates the perturbation amount in the feature space through the Adam optimizer until the decision result is reversed and the perturbation amplitude meets the constraint condition.

[0072] When the perturbation amount that meets the condition is found in the feature space, the counterfactual explanation generation module will use a pre-trained generative adversarial network (GAN) as the image generator . This generator receives the perturbed feature tensor as input and decodes it to reconstruct the counterfactual image variant The adversarial generative network can learn a complex mapping relationship from the feature space to the real image pixel distribution, so the generated counterfactual image is usually more natural, realistic and semantically coherent in vision, avoiding artifacts and unnaturalness that may be caused by direct pixel perturbation. During training, the generator usually forms an autoencoder structure with the encoder part of the multi-scale feature extraction module to ensure that it can effectively reconstruct the image from the features.

[0073] The present embodiment solves the challenge of direct pixel perturbation in generating natural counterfactual explanations by perturbing in the feature space and combining an adversarial generative network for image reconstruction. Perturbing in the feature space allows the system to manipulate image content at a higher semantic level, such as accurately "removing" or "modifying" a certain semantic concept in the image, rather than simply changing the pixel color. Combined with the image reconstruction capability of the adversarial generative network, the generated counterfactual image variants not only successfully change the model's decision, but also maintain a high degree of realism and semantic coherence in vision, making it easier for experts to understand and trust these explanations. For example, in an autonomous driving scenario, a counterfactual image "missing a certain traffic sign" can be generated through feature space perturbation, and this image is still very natural in vision, helping to understand the model's dependence on traffic signs. This greatly improves the quality, understandability and practicality 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 corresponding to adding, deleting, or reversing a causal edge in the causal graph; the agent learns the optimal causal structure by maximizing the reward function, which is jointly 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