An artificial intelligence driven image recognition system and method of recognition thereof

Abstract

The application relates to the technical field of image recognition, in particular to an artificial intelligence driven image recognition system and a recognition method thereof, which comprises an image acquisition module, a preprocessing module, a feature extraction module, an image recognition module and a result output module; the image acquisition module acquires input images or video frames in real time; the preprocessing module performs normalization, noise reduction, enhancement and other operations on the collected image data; the feature extraction module encodes and represents key regions of the image through a multilayer neural structure based on a deep convolutional neural network, and extracts multiscale spatial features; the image recognition module fuses recognition results of multiple neural network models based on an ensemble learning mechanism, and outputs final classification or recognition information; and the result output module outputs the calculated result.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to an artificial intelligence-driven image recognition system and its recognition method. Background Technology

[0002] With the rapid development of computer vision and artificial intelligence technologies, image recognition, as a core branch of artificial intelligence, has been widely applied in various industries such as security monitoring, medical diagnosis, intelligent transportation, and industrial inspection. Traditional image recognition methods mainly rely on manually designed feature extractors, which have certain limitations when dealing with image deformation, noise interference, or complex background environments. In recent years, the rise of deep learning, especially convolutional neural networks, has greatly improved the accuracy and generalization ability of image recognition. Through multi-layer feature learning, semantic information at different levels in images can be effectively extracted, thereby achieving efficient classification and recognition of image targets. However, most existing systems focus on a single neural network model, lacking flexibility and adaptability, and still have biases in the recognition of edge samples and multi-class confusion regions, failing to meet the needs of diverse and complex real-world scenarios. In addition, the computational resource consumption of large models also limits their deployment in resource-constrained environments such as edge devices and mobile devices.

[0003] Therefore, there is an urgent need for a new image recognition system that can not only integrate the advantages of multiple neural network models to improve recognition accuracy while possessing stronger robustness and versatility, but also support the optimization of discrimination for low-quality images, complex backgrounds, and boundary samples. Meanwhile, with increasing user awareness of privacy protection, models trained on traditional centralized data are prone to data leakage risks, which has driven research into privacy-preserving intelligent recognition technologies such as federated learning and adaptive update mechanisms. To achieve the evolution of image recognition systems from "static judgment" to "intelligent adaptation," the system also needs to possess online self-learning capabilities, that is, continuously optimizing model parameters through feedback training to improve long-term performance stability in specific scenarios. Therefore, designing an AI-driven image recognition system that integrates multiple network structures, possesses high robustness, adaptability, and privacy protection mechanisms has become a current technological development trend and research focus in the field of intelligent vision.

[0004] In view of the above, in order to overcome the above technical problems, the present invention designs an artificial intelligence-driven image recognition system and recognition method, which solves the above technical problems. Summary of the Invention

[0005] The technical objective of this invention is to design an artificial intelligence-driven image recognition system and its recognition method, which improves long-term performance stability in specific scenarios by continuously training and optimizing model parameters through feedback.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution: An AI-driven image recognition system is designed to improve the accuracy, robustness, and adaptability of image recognition, making it particularly suitable for applications involving mixed images, blurred boundaries, or poor image quality. The system mainly comprises an image acquisition module, a preprocessing module, a feature extraction module, an image recognition module, a result output module, and a model adaptation module. These modules work collaboratively to form a complete intelligent image recognition process.

[0007] The image acquisition module is used to acquire input image or video frame data in real time, supporting various image sources, including industrial cameras, webcams, mobile terminals, or drones. This module has a certain degree of image stabilization and adaptive exposure control capabilities, ensuring the stability of image input quality. The acquired raw image data is sent to the preprocessing module for image standardization. The preprocessing module performs normalization, noise reduction, and enhancement operations on the image. The normalization step maps image pixel values ​​to a fixed range to reduce the impact of lighting differences; noise reduction removes interference noise from the image, improving the effectiveness of subsequent feature extraction; enhancement processing includes contrast adjustment and edge enhancement to enhance image detail.

[0008] The feature extraction module is one of the core modules of the system, employing a multi-layer neural structure based on a deep convolutional neural network (CNN). This structure can extract multi-scale spatial features from images, learning various information such as edges, textures, shapes, and high-level semantics layer by layer. By introducing residual connections or attention mechanisms into the network, the system can perform more precise representation encoding of key regions, providing rich and accurate feature support for subsequent recognition.

[0009] The image recognition module is based on an ensemble learning mechanism, fusing the recognition results of multiple neural network models to output more robust classification or recognition information. Specifically, the image recognition module includes a main discriminant network, an auxiliary correction network, and a confidence evaluation unit. The main discriminant network is responsible for preliminary classification prediction, using a normalized exponential function (Softmax) to probabilistically output the image feature vector; the auxiliary correction network mainly targets image samples with blurred boundaries or low confidence, using a neural network model based on residual structures to correct and optimize the preliminary results; the confidence evaluation unit, based on the output results of multiple models, corrects the final recognition result through a weighted voting mechanism or a confidence fusion algorithm, improving the overall stability and classification reliability of the system.

[0010] In addition, the system also includes a model adaptation module, which dynamically adjusts the recognition strategy based on the real-time distribution characteristics of the input image. This module analyzes the class weights and sample distribution of the image input, automatically adjusts the discrimination threshold, and performs targeted model parameter optimization and structural adjustments for boundary samples or low-confidence classes, thereby enhancing the system's robustness and generalization ability.

[0011] The results output module is responsible for displaying or outputting the final recognition results to the terminal system. This may include textual label output, coordinate position feedback, or image segmentation result presentation. It supports access to host computer systems, databases, or other application platforms.

[0012] The preprocessing module includes an image normalization unit, a histogram equalization unit, and an edge enhancement unit. These three units work together to optimize the quality of the original image and enhance key features, providing a solid input foundation for subsequent deep neural network processing. The image normalization unit linearly maps the pixel values ​​of the input image to the range [0,1], reducing dynamic differences caused by variations in lighting, uneven exposure, or different imaging devices, thus improving the model's generalization ability. The histogram equalization unit enhances the image's contrast based on its grayscale distribution characteristics. By adjusting the distribution of pixel values, it makes the details in both dark and bright areas clearer, facilitating the subsequent extraction of more discriminative mid-to-high-level semantic features by the subsequent network. The edge enhancement unit uses the Laplacian operator to extract image edges, strengthening contour and structural information in the image through the response to the second derivative, effectively highlighting object boundaries and key contours. Specifically, the edge-enhanced image is input as a feature channel into the first layer of the deep convolutional neural network, participating in the initial convolution operation together with the original image channel. This enhances the model's ability to perceive local details and edge information, helping to improve recognition performance in complex backgrounds or low-contrast scenes.

[0013] The main discriminator network in the image recognition module employs three mainstream deep neural network structures—Visual Geometric Group Network (VGG), Residual Neural Network (ResNet), and EfficientNet—running in parallel. This multi-model collaboration enhances recognition accuracy and robustness. Each network independently receives the same image feature input and performs classification prediction. The system dynamically allocates the weights of each network in the final result fusion based on their cross-validation accuracy during the training phase. Specifically, an ensemble weighted average strategy is used, weighting and superimposing the probability distributions of the three models to obtain a more representative and stable classification probability. The final recognition output selects the class label with the highest weighted probability value as the system's recognition result, effectively mitigating the problem of misjudging edge samples by a single model and improving the overall system's adaptability and judgment ability in complex image environments.

[0014] The image recognition results obtained by the output module have broad application adaptability and can be directly applied to various practical scenarios such as medical image recognition (e.g., tumor detection, lesion localization), intelligent traffic monitoring (e.g., license plate recognition, violation analysis), and facial recognition (e.g., identity verification and behavior tracking). The system supports loading specialized datasets built for different industries and combining them with transfer learning techniques to fine-tune the model, thereby quickly adapting to the recognition requirements of specific tasks. Simultaneously, this module integrates a multi-task learning mechanism, executing object detection, image segmentation, and image classification tasks in parallel within a unified deep neural network structure. This not only improves overall computational efficiency but also significantly enhances the system's multi-functional recognition capabilities and understanding of complex image scenes, providing a highly flexible and scalable image processing solution for practical deployment.

[0015] The auxiliary correction network uses an attention mechanism to assign weights to intermediate feature maps, highlighting the salient features of the target region. This attention mechanism is defined as follows: Where Q, K, and V are query, key, and value vectors, and dk is the dimension; this mechanism enhances the ability to distinguish boundary samples.

[0016] The confidence assessment unit outputs a confidence score for each recognition result. The confidence score is calculated as follows: Where zi is the output value of the i-th class. If the confidence scores of all classes are less than the set threshold T, the manual review interface is triggered.

[0017] The model adaptation module consists of a classifier dynamic adjustment unit and a training data resampling unit, designed to address class imbalance and overfitting issues. The classifier dynamic adjustment unit introduces a temperature scaling algorithm during the inference phase to adjust the output of the normalized exponential function (Softmax). By adjusting the temperature coefficient T, it reduces the model's overconfidence in the dominant class, thereby improving its ability to distinguish long-tail class samples and mitigating prediction bias. Simultaneously, the training data resampling unit analyzes the density distribution of samples in the feature space, upsampling low-frequency samples and downsampling or weighting high-frequency samples, automatically generating new training batches to improve the model's generalization ability on multi-class data and effectively avoid overfitting. This module supports dynamic linkage between the training and inference phases, enhancing the system's adaptability and stability in the face of changing data distributions in practical applications.

[0018] The system is deployed entirely on edge computing devices to fully leverage the advantages of edge computing in terms of real-time performance and data security. Specifically, the image acquisition and preprocessing modules are integrated into the front-end acquisition device, enabling real-time image acquisition and preliminary preprocessing, such as normalization, noise reduction, and enhancement, to be performed directly on the device, reducing data transmission volume and improving data quality. The feature extraction and image recognition modules are deployed on the edge computing unit using lightweight neural network models. These models are optimized using pruning and quantization techniques, resulting in lower computational resource requirements and ensuring efficient system operation even with limited hardware. This architecture supports autonomous recognition in offline environments while achieving low-latency real-time image processing, meeting the stringent requirements of rapid response and local processing in applications such as intelligent monitoring, autonomous driving, and mobile devices. The distributed deployment of edge computing also effectively protects user data privacy and security, avoiding the uploading of large amounts of sensitive image data to the cloud, thus enhancing the system's practicality and reliability.

[0019] The system supports a federated learning-based model update mechanism to achieve distributed collaborative training and protect user privacy. Specifically, each terminal device independently trains its model locally, updating model parameters using its own collected image data, and uploading the gradient information or model weight differences calculated during training to a central server. The central server securely aggregates the gradient information from each terminal, generates the latest parameters of the global model, and then broadcasts this global model back to all terminal devices, achieving unified model updates. This process eliminates the need to upload original image data, effectively avoiding the risk of user privacy leaks and complying with data protection regulations. Simultaneously, the federated learning mechanism significantly improves the model's generalization ability and adaptability by integrating diverse data features from multiple terminals, ensuring the system maintains high recognition accuracy across various application scenarios. This mechanism is particularly suitable for applications in data-scarce and highly sensitive environments, such as medical image recognition and intelligent monitoring, balancing the dual needs of privacy protection and model performance.

[0020] An artificial intelligence-driven image recognition method, used in conjunction with the aforementioned artificial intelligence-driven image recognition system; characterized in that the method comprises the following steps: S1: Acquires raw image data or video frame data through the image acquisition module, supporting static images and dynamic frame input, and image formats include standard formats such as JPEG, PNG or YUV; S2: Normalize, equalize, and enhance the edges of the acquired image. The normalization method is to linearly map to the [0,1] interval, and the enhancement method includes the Laplacian operator. S3: Use a deep convolutional neural network to perform multi-layer convolution and pooling operations on the preprocessed image to extract multi-scale semantic features. The convolutional structure uses residual connections, and the output of the intermediate layer of the network is used as the input feature φ for subsequent recognition. S4: The extracted features are input into the main discriminant network and the auxiliary correction network for initial judgment and correction, respectively. The main network uses softmax for classification prediction, and the auxiliary network uses an attention mechanism to optimize boundary samples. Finally, the class label with the highest confidence is output by ensemble weighted average. S5: Output the final recognition result and its confidence level. If the confidence level is lower than the preset threshold T, the review mechanism is triggered. At the same time, the input sample features are recorded for adaptive module training to dynamically update the model weights and achieve continuous optimization of recognition accuracy.

[0021] The beneficial effects of this invention are as follows: (1) This invention effectively improves the accuracy and robustness of image classification and recognition by introducing a multi-network structure parallel fusion image recognition system. It employs three mainstream deep neural networks—Visual Geometric Group Network (VGG), Residual Neural Network (ResNet), and EfficientNet—running in parallel, combined with a dynamic weighted fusion strategy. This achieves the complementarity of the advantages of different networks, significantly reducing the misjudgment and bias problems that are prone to occur with single models. Simultaneously, the design of the auxiliary correction network and confidence evaluation unit further optimizes the discrimination ability for boundary samples and low-confidence results, making the system more stable and adaptable when processing complex and diverse images. The system's multi-level, multi-angle feature extraction and discrimination method greatly improves the ability to capture and express image details, effectively enhancing the accuracy and reliability of recognition, and meeting the demand for high-performance image recognition in practical applications.

[0022] (2) This invention also possesses significant advantages in adaptability and privacy protection. The model adaptation module enables dynamic adjustment of category distribution and sample structure, enhancing the recognition of long-tailed classes and marginal samples, avoiding overfitting, and improving the model's generalization ability and stability. Combined with the federated learning mechanism, this system supports local training on terminal devices and model aggregation updates on the central server, effectively protecting user privacy data security and avoiding the risk of sensitive information leakage. This distributed training architecture not only improves the system's real-time update capability but also adapts to the diverse needs of different terminals and scenarios, greatly enhancing the system's practicality and promotional value. In summary, this invention provides a high-performance, flexible, and secure AI-driven image recognition solution with broad application prospects and significant industrial promotion value. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] The above and other aspects of the invention will now be described by way of example only, with reference to the accompanying drawings, in which: Figure 1 This is a schematic diagram of the system flow of the present invention; Figure 2 This is a flowchart illustrating the steps of the method of the present invention. Detailed Implementation

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] like Figure 1-2 As shown, an AI-driven image recognition system aims to improve the accuracy, robustness, and adaptability of image recognition, making it particularly suitable for application scenarios involving mixed images, blurred boundaries, or poor image quality. The system mainly includes an image acquisition module, a preprocessing module, a feature extraction module, an image recognition module, a result output module, and a model adaptation module. These modules work together to form a complete intelligent image recognition process.

[0027] The image acquisition module is used to acquire input image or video frame data in real time, supporting various image sources, including industrial cameras, webcams, mobile terminals, or drones. This module has a certain degree of image stabilization and adaptive exposure control capabilities, ensuring the stability of image input quality. The acquired raw image data is sent to the preprocessing module for image standardization. The preprocessing module performs normalization, noise reduction, and enhancement operations on the image. The normalization step maps image pixel values ​​to a fixed range to reduce the impact of lighting differences; noise reduction removes interference noise from the image, improving the effectiveness of subsequent feature extraction; enhancement processing includes contrast adjustment and edge enhancement to enhance image detail.

[0028] The feature extraction module is one of the core modules of the system, employing a multi-layer neural structure based on a deep convolutional neural network (CNN). This structure can extract multi-scale spatial features from images, learning various information such as edges, textures, shapes, and high-level semantics layer by layer. By introducing residual connections or attention mechanisms into the network, the system can perform more precise representation encoding of key regions, providing rich and accurate feature support for subsequent recognition.

[0029] The image recognition module is based on an ensemble learning mechanism, fusing the recognition results of multiple neural network models to output more robust classification or recognition information. Specifically, the image recognition module includes a main discriminant network, an auxiliary correction network, and a confidence evaluation unit. The main discriminant network is responsible for preliminary classification prediction, using a normalized exponential function (Softmax) to probabilistically output the image feature vector; the auxiliary correction network mainly targets image samples with blurred boundaries or low confidence, using a neural network model based on residual structures to correct and optimize the preliminary results; the confidence evaluation unit, based on the output results of multiple models, corrects the final recognition result through a weighted voting mechanism or a confidence fusion algorithm, improving the overall stability and classification reliability of the system.

[0030] In addition, the system also includes a model adaptation module, which dynamically adjusts the recognition strategy based on the real-time distribution characteristics of the input image. This module analyzes the class weights and sample distribution of the image input, automatically adjusts the discrimination threshold, and performs targeted model parameter optimization and structural adjustments for boundary samples or low-confidence classes, thereby enhancing the system's robustness and generalization ability.

[0031] The results output module is responsible for displaying or outputting the final recognition results to the terminal system. This may include textual label output, coordinate position feedback, or image segmentation result presentation. It supports access to host computer systems, databases, or other application platforms.

[0032] The preprocessing module includes an image normalization unit, a histogram equalization unit, and an edge enhancement unit. These three units work together to optimize the quality of the original image and enhance key features, providing a solid input foundation for subsequent deep neural network processing. The image normalization unit linearly maps the pixel values ​​of the input image to the range [0,1], reducing dynamic differences caused by variations in lighting, uneven exposure, or different imaging devices, thus improving the model's generalization ability. The histogram equalization unit enhances the image's contrast based on its grayscale distribution characteristics. By adjusting the distribution of pixel values, it makes the details in both dark and bright areas clearer, facilitating the subsequent extraction of more discriminative mid-to-high-level semantic features by the subsequent network. The edge enhancement unit uses the Laplacian operator to extract image edges, strengthening contour and structural information in the image through the response to the second derivative, effectively highlighting object boundaries and key contours. Specifically, the edge-enhanced image is input as a feature channel into the first layer of the deep convolutional neural network, participating in the initial convolution operation together with the original image channel. This enhances the model's ability to perceive local details and edge information, helping to improve recognition performance in complex backgrounds or low-contrast scenes.

[0033] The main discriminator network in the image recognition module employs three mainstream deep neural network structures—Visual Geometric Group Network (VGG), Residual Neural Network (ResNet), and EfficientNet—running in parallel. This multi-model collaboration enhances recognition accuracy and robustness. Each network independently receives the same image feature input and performs classification prediction. The system dynamically allocates the weights of each network in the final result fusion based on their cross-validation accuracy during the training phase. Specifically, an ensemble weighted average strategy is used, weighting and superimposing the probability distributions of the three models to obtain a more representative and stable classification probability. The final recognition output selects the class label with the highest weighted probability value as the system's recognition result, effectively mitigating the problem of misjudging edge samples by a single model and improving the overall system's adaptability and judgment ability in complex image environments.

[0034] The image recognition results obtained by the output module have broad application adaptability and can be directly applied to various practical scenarios such as medical image recognition (e.g., tumor detection, lesion localization), intelligent traffic monitoring (e.g., license plate recognition, violation analysis), and facial recognition (e.g., identity verification and behavior tracking). The system supports loading specialized datasets built for different industries and combining them with transfer learning techniques to fine-tune the model, thereby quickly adapting to the recognition requirements of specific tasks. Simultaneously, this module integrates a multi-task learning mechanism, executing object detection, image segmentation, and image classification tasks in parallel within a unified deep neural network structure. This not only improves overall computational efficiency but also significantly enhances the system's multi-functional recognition capabilities and understanding of complex image scenes, providing a highly flexible and scalable image processing solution for practical deployment.

[0035] The auxiliary correction network uses an attention mechanism to assign weights to intermediate feature maps, highlighting the salient features of the target region. This attention mechanism is defined as follows: Where Q, K, and V are query, key, and value vectors, and dk is the dimension; this mechanism enhances the ability to distinguish boundary samples.

[0036] The confidence assessment unit outputs a confidence score for each recognition result. The confidence score is calculated as follows: Where zi is the output value of the i-th class. If the confidence scores of all classes are less than the set threshold T, the manual review interface is triggered.

[0037] The model adaptation module consists of a classifier dynamic adjustment unit and a training data resampling unit, designed to address class imbalance and overfitting issues. The classifier dynamic adjustment unit introduces a temperature scaling algorithm during the inference phase to adjust the output of the normalized exponential function (Softmax). By adjusting the temperature coefficient T, it reduces the model's overconfidence in the dominant class, thereby improving its ability to distinguish long-tail class samples and mitigating prediction bias. Simultaneously, the training data resampling unit analyzes the density distribution of samples in the feature space, upsampling low-frequency samples and downsampling or weighting high-frequency samples, automatically generating new training batches to improve the model's generalization ability on multi-class data and effectively avoid overfitting. This module supports dynamic linkage between the training and inference phases, enhancing the system's adaptability and stability in the face of changing data distributions in practical applications.

[0038] The system is deployed entirely on edge computing devices to fully leverage the advantages of edge computing in terms of real-time performance and data security. Specifically, the image acquisition and preprocessing modules are integrated into the front-end acquisition device, enabling real-time image acquisition and preliminary preprocessing, such as normalization, noise reduction, and enhancement, to be performed directly on the device, reducing data transmission volume and improving data quality. The feature extraction and image recognition modules are deployed on the edge computing unit using lightweight neural network models. These models are optimized using pruning and quantization techniques, resulting in lower computational resource requirements and ensuring efficient system operation even with limited hardware. This architecture supports autonomous recognition in offline environments while achieving low-latency real-time image processing, meeting the stringent requirements of rapid response and local processing in applications such as intelligent monitoring, autonomous driving, and mobile devices. The distributed deployment of edge computing also effectively protects user data privacy and security, avoiding the uploading of large amounts of sensitive image data to the cloud, thus enhancing the system's practicality and reliability.

[0039] The system supports a federated learning-based model update mechanism to achieve distributed collaborative training and protect user privacy. Specifically, each terminal device independently trains its model locally, updating model parameters using its own collected image data, and uploading the gradient information or model weight differences calculated during training to a central server. The central server securely aggregates the gradient information from each terminal, generates the latest parameters of the global model, and then broadcasts this global model back to all terminal devices, achieving unified model updates. This process eliminates the need to upload original image data, effectively avoiding the risk of user privacy leaks and complying with data protection regulations. Simultaneously, the federated learning mechanism significantly improves the model's generalization ability and adaptability by integrating diverse data features from multiple terminals, ensuring the system maintains high recognition accuracy across various application scenarios. This mechanism is particularly suitable for applications in data-scarce and highly sensitive environments, such as medical image recognition and intelligent monitoring, balancing the dual needs of privacy protection and model performance.

[0040] An artificial intelligence-driven image recognition method, used in conjunction with the aforementioned artificial intelligence-driven image recognition system; characterized in that the method comprises the following steps: S1: Acquires raw image data or video frame data through the image acquisition module, supporting static images and dynamic frame input, and image formats include standard formats such as JPEG, PNG or YUV; S2: Normalize, equalize, and enhance the edges of the acquired image. The normalization method is to linearly map to the [0,1] interval, and the enhancement method includes the Laplacian operator. S3: Use a deep convolutional neural network to perform multi-layer convolution and pooling operations on the preprocessed image to extract multi-scale semantic features. The convolutional structure uses residual connections, and the output of the intermediate layer of the network is used as the input feature φ for subsequent recognition. S4: The extracted features are input into the main discriminant network and the auxiliary correction network for initial judgment and correction, respectively. The main network uses softmax for classification prediction, and the auxiliary network uses an attention mechanism to optimize boundary samples. Finally, the class label with the highest confidence is output by ensemble weighted average. S5: Output the final recognition result and its confidence level. If the confidence level is lower than the preset threshold T, the review mechanism is triggered. At the same time, the input sample features are recorded for adaptive module training to dynamically update the model weights and achieve continuous optimization of recognition accuracy.

[0041] Various modifications to this disclosure will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but should be given the broadest scope consistent with the principles and novel features disclosed herein. Although one or more exemplary embodiments of this disclosure have been described with reference to the accompanying drawings, those skilled in the art will understand that various changes in form and detail may be made therein without departing from the spirit and scope of this disclosure as defined by the appended claims.

Claims

1. An artificial intelligence-driven image recognition system, characterized in that, It includes an image acquisition module, a preprocessing module, a feature extraction module, an image recognition module, and a result output module; Image acquisition module: Real-time acquisition of input images or video frames; The preprocessing module performs normalization, noise reduction, and enhancement operations on the acquired image data; Feature extraction module: Encodes and represents key regions of the image using a multi-layer neural structure based on a deep convolutional neural network, and extracts multi-scale spatial features; The image recognition module integrates the recognition results of multiple neural network models based on an ensemble learning mechanism, and outputs the final classification or recognition information. The image recognition module specifically includes: a main discriminant network, an auxiliary correction network, and a confidence evaluation unit. The main discriminant network uses a normalized exponential function for initial classification prediction, and the auxiliary correction network uses a residual structure for secondary judgment of edge samples. The auxiliary correction network shares the multi-scale spatial features output by the feature extraction module with the main discriminant network. The auxiliary correction network uses an attention mechanism to assign weights to intermediate feature maps, highlighting the salient features of the target region. The confidence evaluation unit corrects the final output result based on a weighted voting mechanism. The system further includes a model adaptation module, which dynamically adjusts the recognition threshold according to the category distribution of the input image to improve the recognition accuracy and stability of boundary samples. The model adaptation module includes a classifier dynamic adjustment unit and a training data resampling unit. The dynamic adjustment unit uses a temperature scaling algorithm to correct the output of the normalized exponential function according to the category distribution, and the resampling unit generates new training batches according to the sample density. The two are dynamically linked during the training and inference stages. Result output module: Outputs the calculated results; The image acquisition module has certain image stabilization and adaptive exposure control capabilities; the enhancement processing includes contrast adjustment, edge enhancement operations, and enhancement of image detail information. The preprocessing module includes an image normalization unit, a histogram equalization unit, and an edge enhancement unit. The image normalization unit maps pixel values ​​to the range of [0,1]. The equalization unit enhances contrast based on the grayscale distribution of the image. The edge enhancement unit uses the Laplacian operator to extract edge features. The edge-enhanced image is then used as a feature channel input to the first layer of the deep convolutional neural network, where it participates in the initial convolution operation along with the original image channels.

2. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The main discriminant network in the image recognition module adopts three structures that run in parallel: visual geometric group network, residual neural network and efficient convolutional neural network. Its classification output is fused by ensemble weighted averaging. The fusion weight is dynamically adjusted according to the cross-validation accuracy during training. The final recognition output is the class with the highest probability of weighted voting.

3. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The results obtained in the output module can be applied to multiple scenarios such as medical image recognition, traffic monitoring, and face recognition. By configuring specific datasets for domain fine-tuning, it supports a multi-task learning mechanism, integrating target detection, image segmentation, and classification tasks into a unified network structure, thereby improving the system's multi-functional recognition capabilities.

4. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The attention mechanism is defined as follows: Where Q, K, and V are query, key, and value vectors, respectively, and d k This is a dimension; the mechanism enhances the discrimination ability of boundary samples.

5. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The confidence assessment unit outputs a confidence score for each recognition result. The confidence score is calculated as follows: Where z i The output value for the i-th class is given. If the confidence scores for all classes are less than the set threshold, the manual review interface is triggered.

6. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The system is deployed on edge computing devices. The image acquisition module and preprocessing module are embedded in the front-end device, while the feature extraction and recognition module is deployed on the edge computing unit through a lightweight model, supporting offline recognition and low-latency real-time processing.

7. The artificial intelligence-driven image recognition system according to claim 1, characterized in that: The system supports a federated learning-based model update mechanism, where each terminal device trains model parameters locally and uploads gradient information to the central server. The central server then aggregates the data and broadcasts the global model. This mechanism protects user privacy and ensures the model's generalization ability.

8. An artificial intelligence-driven image recognition method, used in conjunction with an artificial intelligence-driven image recognition system according to any one of claims 1-7; characterized in that: The steps of the method are as follows: S1: Acquires raw image data or video frame data through the image acquisition module, supporting static images and dynamic frame input, with image formats including JPEG, PNG or YUV standard formats; S2: Normalize, equalize, and enhance the edges of the acquired image. The normalization method is to linearly map to the [0,1] interval, and the enhancement method includes the Laplacian operator. S3: Use a deep convolutional neural network to perform multi-layer convolution and pooling operations on the preprocessed image to extract multi-scale semantic features. The convolutional structure uses residual connections, and the output of the intermediate layer of the network is used as the input feature φ for subsequent recognition. S4: The extracted features are input into the main discriminant network and the auxiliary correction network for initial judgment and correction, respectively. The main network uses softmax for classification prediction, and the auxiliary network uses an attention mechanism to optimize boundary samples. Finally, the class label with the highest confidence is output by ensemble weighted average. S5: Output the final recognition result and its confidence level. If the confidence level is lower than the preset threshold, trigger the review mechanism. At the same time, record the input sample features for adaptive module training to dynamically update the model weights and achieve continuous optimization of recognition accuracy.

Images