Image classification system and method based on image recognition technology

Abstract

The invention relates to the technical field of image recognition, in particular to an image classification system and method based on the image recognition technology, and the system comprises an image collection module which is used for obtaining original image data to be classified; the preprocessing module is used for carrying out denoising, normalization and size standardization processing on the image; the feature extraction module is used for extracting multi-level features of the image by adopting a deep convolutional neural network; the classification decision module is used for weighting fusion features based on an attention mechanism and outputting a classification result; the output module is used for displaying the classification labels and confidence scores; according to the method, the input quality is optimized by dynamically selecting a preprocessing strategy, the multi-scale representation capability is enhanced by adopting a parallel convolution path and a feature pyramid structure, the robustness of the system is improved by integrating an adversarial sample detection and defense mechanism, and the dynamic scheduling and mixing precision acceleration of computing resources are realized by introducing an edge computing optimization technology. And the operation efficiency is obviously improved on the premise of ensuring the classification precision.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to an image classification system and method based on image recognition technology. Background Art

[0002] With the rapid development of computer vision technology, image classification systems have been widely used in many fields, such as medical image analysis, autonomous driving, and industrial quality inspection. Existing image classification technologies mainly rely on deep convolutional neural networks for feature extraction, combined with classifiers such as fully connected layers or support vector machines to achieve category prediction. Some improvement schemes introduce attention mechanisms or multi-scale feature fusion strategies to improve classification accuracy. At the same time, some studies have also attempted to optimize system performance through model lightweighting or adversarial training. However, these existing technologies still face many challenges in actual deployment.

[0003] Current image classification systems have many shortcomings. First, traditional preprocessing methods usually adopt fixed denoising and enhancement strategies, which are difficult to adapt to fluctuations in image quality under different imaging conditions, resulting in a significant decrease in classification performance in low-light or high-noise scenarios; second, existing feature extraction networks often only focus on single-scale features, lack the effective coordinated utilization of fine-grained local features and global semantic information, and perform poorly when faced with fine classification tasks with small inter-class differences; third, most systems have insufficient defense capabilities against adversarial sample attacks, and even slight perturbations can lead to misclassification, which seriously restricts the reliability of applications in security-sensitive fields; fourth, existing technologies generally lack dynamic resource allocation mechanisms, and use the same computing load when processing samples of different complexities, resulting in low energy efficiency on edge devices.

[0004] To address the above problems, the present invention proposes an image classification system and method based on image recognition technology. The system optimizes input quality by dynamically selecting preprocessing strategies, adopts parallel convolution paths and feature pyramid structures to enhance multi-scale characterization capabilities, and integrates adversarial sample detection and defense mechanisms to improve system robustness. At the same time, it introduces edge computing optimization technology to achieve dynamic scheduling of computing resources and mixed-precision acceleration, significantly improving operational efficiency while ensuring classification accuracy. Summary of the Invention

[0005] In order to overcome the problem that traditional preprocessing methods generally adopt fixed denoising and enhancement strategies, which are difficult to adapt to fluctuations in image quality under different imaging conditions, resulting in a significant decrease in classification performance in low-light or high-noise scenes, the present invention proposes an image classification system and method based on image recognition technology.

[0006] The technical solution of the present invention is: an image classification system and method based on image recognition technology, including:

[0007] Image acquisition module: used to obtain the original image data to be classified;

[0008] Preprocessing module: denoising, normalizing and size standardizing the image;

[0009] Feature extraction module: uses deep convolutional neural network to extract multi-level features of images;

[0010] Classification decision module: weighted fusion features based on the attention mechanism and output classification results;

[0011] Output module: displays classification labels and confidence scores.

[0012] Preferably, the preprocessing module dynamically selects a denoising algorithm based on image quality assessment, selects a denoising strategy by calculating the signal-to-noise ratio of the image block, and adaptively corrects brightness, adjusting low-contrast images through histogram equalization.

[0013] Preferably, the feature extraction module uses a multi-scale feature extraction network, which includes a parallel convolution path and a feature pyramid structure. The parallel convolution path is used to extract local features of different receptive fields. The first path is a 3×3 convolution kernel with a step size of 1 and an output feature map size of H×W×64. The second path is a 5×5 void convolution, which outputs a feature map of the same size; the feature pyramid structure is used to fuse shallow texture features with deep semantic features.

[0014] Preferably, the classification decision module uses an explainable classification method for the generation of heat maps and the output of multi-level classification. The generation of heat maps is based on the Grad-CAM++ algorithm to calculate category-sensitive areas, and the resolution of the heat map maintains a 1:4 ratio with the input image, and Jet color mapping is used for superimposed display. The output of multi-level classification is divided into three levels: the primary classification result outputs the Top-1 classification label and its confidence score; the secondary alternative result, when the Top-1 confidence is <0.7, the Top-3 categories and their probability distribution are output; uncertainty annotation, samples with a confidence of <0.5 are marked as "low reliability".

[0015] Preferably, the system further comprises a new category learning unit and a sample memory library. The new category learning unit realizes online model update by freezing the underlying network layer and fine-tuning the classifier; the sample memory library includes a storage strategy, dynamic update and playback mechanism;

[0016] Storage strategy: Select representative samples based on feature space K-means clustering;

[0017] Dynamic update: When the distance between a new sample and an existing cluster center is greater than the threshold, it is stored in the memory bank;

[0018] Replay mechanism: During training, new data and memory samples are mixed in a 1:1 ratio.

[0019] Preferably, the image classification method based on image recognition technology includes the following steps:

[0020] S1, first obtains the raw image data to be classified through the image acquisition module, and then enters the preprocessing module for automated processing. This module first evaluates the image quality to determine the appropriate denoising algorithm. For low-quality images, non-local mean filtering is used, while for high-quality images, fast bilateral filtering is used. At the same time, adaptive brightness correction is performed through CLAHE algorithm partition processing and combined with nonlinear brightness compression function to adjust the overexposed area. Finally, geometric correction is performed to eliminate image tilt and perspective distortion to ensure that the input image meets the standard requirements of subsequent feature extraction;

[0021] S2, the preprocessed image is input into the hybrid neural network architecture of the feature extraction module, which consists of a combination of a ResNet-50 backbone network and an inserted Transformer layer. First, a parallel convolution path is used to extract local features of different receptive fields using 3×3 standard convolution and 5×5 dilated convolution respectively. Then, shallow texture features and deep semantic features are fused from top to bottom in the feature pyramid structure, and finally a multi-scale feature map with four levels from P2 to P5 is output, where high-level features contain rich semantic information while low-level features retain fine spatial details.

[0022] S3, the dual attention mechanism that inputs the extracted multi-scale features into the classification decision module, first recalibrates the importance of the channel dimension of the feature map through the channel attention mechanism to suppress the response of irrelevant channels, then uses the spatial attention mechanism to focus on the key areas related to classification in the spatial dimension, and generates a spatial attention weight map through cascaded global average pooling and maximum pooling operations. Finally, the weighted feature map is input into the subsequent classifier;

[0023] In S4, the attention-weighted features are first compressed in spatial dimensions through a global average pooling layer, and then input into a classifier containing a fully connected layer for category prediction. The output is calibrated using a temperature-adjusted Softmax function to balance the confidence distribution of each category. At the same time, the cross-entropy loss is calculated for parameter optimization in the model training phase. In the inference phase, not only the Top-1 classification label and its confidence score are output, but also the Top-3 candidate categories and their probability distribution are provided when the confidence of the main classification result is lower than the threshold. A visual heat map is generated through the Grad-CAM++ algorithm to enhance the interpretability of the classification results.

[0024] Preferably, after step S1 is completed, the input image is fast Fourier transformed and the high-frequency component features are extracted. When an abnormal high-frequency energy distribution is detected, an adversarial sample alarm is triggered, and then a pixel-level defense process is started. This process uses an image reconstructor based on a generative adversarial network to repair local areas of suspected adversarial samples. The reconstructor achieves semantic-level reconstruction of damaged pixels through a pre-trained U-Net architecture. At the same time, randomization processing is introduced in the feature extraction stage, including adding Gaussian noise perturbation to the weights of the convolutional layer and performing random spatial transformation on the feature map. Finally, a confidence threshold detection mechanism is deployed in the classification decision module. When the confidence of the output prediction is lower than the preset threshold, the sample is automatically marked as a suspicious adversarial sample and a secondary verification process is started. The robustness of the final classification result is ensured through a multi-model voting mechanism.

[0025] Preferably, when the input data contains associated text information, the cross-modal joint analysis process is started. First, image features and text features are extracted respectively through independent visual encoders and text encoders. The visual encoder adopts a multi-scale feature extraction network, and the text encoder implements semantic embedding based on the pre-trained BERT model. Subsequently, a visual-semantic association model is constructed through a graph neural network, and image region features and text keywords are mapped to a unified graph structure. At the same time, the graph attention mechanism is used to learn the relationship weights between modalities. Finally, the visual features and text features are weightedly spliced through the cross-modal feature fusion layer and input into the multimodal classifier for joint prediction. The classifier uses bilinear pooling technology to capture high-order interaction features between modalities, and supports two working states: pure visual mode and visual-text hybrid mode. When text information is missing, it automatically degenerates to a standard image classification process.

[0026] Preferably, the method also uses an edge computing deployment solution. First, the knowledge distillation technology is used to compress the original large classification model into a lightweight version, wherein the teacher model adopts a network of an interpretable classification method, and the student model adopts a micro-architecture constructed by deep separable convolution. Efficient knowledge transfer is achieved through feature map matching loss and classification logits distillation loss. The compressed model is then fixed-point quantized, and the 32-bit floating-point parameters are converted into 8-bit integer representation and combined with a dynamic range calibration algorithm to keep the accuracy loss within 1%. During runtime, the system dynamically adjusts the calculation path according to the resource status of the edge device, automatically enables fast inference mode for low-complexity images, skips some network layers, and integrates a hardware acceleration interface to optimize convolution operations for ARM architecture processors and support dedicated instruction set acceleration for GPU / NPU.

[0027] Preferably, dynamic computation paths are enabled during the inference phase, thereby skipping redundant network layers for simple samples, while mixed-precision computation is used to accelerate feature extraction.

[0028] Beneficial effects of the present invention:

[0029] 1. The preprocessing module significantly improves the robustness of image quality under different imaging conditions. Combined with a multi-scale feature fusion network, it achieves a coordinated optimization representation of local details and global semantics, improving classification accuracy in complex scenes.

[0030] 2. An innovative adversarial example defense mechanism effectively resists perturbation attacks through frequency domain detection and pixel-level reconstruction, improving system security.

[0031] 3. The edge computing solution uses knowledge distillation and dynamic computing path selection technology, which can simultaneously achieve the advantages of reducing computing latency and reducing memory usage while maintaining the accuracy of the original model.

[0032] 4. Interpretability of the classification decision module The heat map visualization and multi-level confidence output provided by the classification method greatly enhance the credibility and practicality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a schematic diagram of the system structure of the present invention;

[0034] Figure 2 What is shown is a schematic diagram of the system flow of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0036] See also Figure 1 The present invention provides an embodiment: an image classification system based on image recognition technology, comprising:

[0037] Image acquisition module: used to obtain the original image data to be classified;

[0038] Preprocessing module: denoising, normalizing and size standardizing the image;

[0039] Feature extraction module: uses deep convolutional neural network to extract multi-level features of images;

[0040] Classification decision module: weighted fusion features based on the attention mechanism and output classification results;

[0041] Output module: displays classification labels and confidence scores.

[0042] The preprocessing module dynamically selects a denoising algorithm based on image quality assessment, selects a denoising strategy by calculating the signal-to-noise ratio of an image block, and simultaneously performs adaptive brightness correction and adjusts low-contrast images through histogram equalization.

[0043] The feature extraction module uses a multi-scale feature extraction network, which includes parallel convolution paths and feature pyramid structures. The parallel convolution paths are used to extract local features of different receptive fields. The first path is a 3×3 convolution kernel with a step size of 1 and outputs a feature map of size H×W×64. The second path is a 5×5 dilated convolution that outputs a feature map of the same size. The feature pyramid structure is used to fuse shallow texture features with deep semantic features.

[0044] The classification decision module uses an explainable classification method for heat map generation and multi-level classification output. The heat map generation is based on the Grad-CAM++ algorithm to calculate category-sensitive areas, and the resolution of the heat map maintains a 1:4 ratio with the input image. Jet color mapping is used for overlay display. The output of the multi-level classification is divided into three levels: the primary classification result outputs the Top-1 classification label and its confidence score; the secondary alternative result outputs the Top-3 categories and their probability distribution when the Top-1 confidence is less than 0.7; and uncertainty annotation marks samples with a confidence score less than 0.5 as "low reliability."

[0045] The system also includes a new category learning unit and a sample memory library. The new category learning unit realizes online model update by freezing the underlying network layer and fine-tuning the classifier; the sample memory library includes a storage strategy, dynamic update and playback mechanism;

[0046] Storage strategy: Select representative samples based on feature space K-means clustering;

[0047] Dynamic update: When the distance between a new sample and an existing cluster center is greater than the threshold, it is stored in the memory bank;

[0048] Replay mechanism: During training, new data and memory samples are mixed in a 1:1 ratio.

[0049] See also Figure 2 , the image classification method based on image recognition technology, the specific steps are as follows:

[0050] In S1, the system first obtains the raw image data to be classified through the image acquisition module. The system supports multiple input sources such as RGB cameras, infrared sensors, and medical images (such as DICOM), and automatically performs format standardization processing, such as JPEG→Tensor conversion. In view of the possible color gamut differences between different imaging devices, the system has a built-in color space adaptation layer to ensure the consistency of input data. For video stream input, the system adopts a key frame extraction strategy to sample static images at fixed intervals and send them to the subsequent process.

[0051] S2. After the original image enters the preprocessing module, the system first performs an automated quality assessment. By calculating the local signal-to-noise ratio (SNR) and frequency domain energy distribution, the system determines the applicable denoising algorithm. When the SNR is ≤ 15dB, the non-local mean filter is enabled; otherwise, the fast bilateral filter is used. The partitioned CLAHE algorithm is then used to optimize the low-contrast areas, while nonlinear brightness compression is applied to the overexposed areas. Finally, the image geometric distortion is corrected through affine transformation, and the preprocessed image of a standardized size (e.g., 224×224) is output.

[0052] S3, the preprocessed image input feature extraction network, which contains two parallel convolution paths: the first path uses 3×3 standard convolution to extract fine-grained texture features, and the second path uses 5×5 void convolution to capture a wide range of context information. The dual-path feature maps are fused by element-by-element addition and sent to the feature pyramid structure: high-level features are spliced with shallow features through bilinear upsampling to form four-level multi-scale representations of P2 (1 / 4 resolution), P3 (1 / 8), P4 (1 / 16), and P5 (1 / 32). Among them, P2 retains the most spatial details and P5 contains high-level semantic information.

[0053] S4. After feature extraction is completed, the system starts the adversarial sample detection mechanism. First, it performs a fast Fourier transform (FFT) on the feature map to analyze the energy anomalies of the high-frequency components. If adversarial perturbations are detected, the defense process is triggered. This process repairs the damaged pixel area through the pre-trained U-Net reconstruction network, and at the same time applies random Gaussian noise and spatial rotation in the feature space to destroy the adversarial perturbations. The defended image will re-enter the feature extraction stage for secondary classification.

[0054] S5, the verified feature map is input into the classification decision module. The channel attention (SE-Net) first calculates the weight of each channel and suppresses irrelevant feature channels; the spatial attention (CBAM) then generates a spatial weight heat map and focuses on the key areas related to the category. The weighted features are compressed by global average pooling and sent to the fully connected classifier. The classifier uses the Softmax function with temperature adjustment (T=2.0) to output the category probability distribution. The system simultaneously generates a Grad-CAM++ heat map, which visualizes the area based on the decision by calculating the gradient response of the target category to the feature map.

[0055] S6, the final output includes structured classification results and a visualization interface: the main classification label (Top-1) and its confidence (0-1) are the core outputs. When the confidence is <0.7, the Top-3 candidate categories are additionally provided. The visualization interface superimposes the heat map and the original image and marks the key discrimination areas. At the same time, the system continuously monitors the distribution of classification confidence. When low-confidence samples appear continuously, the incremental learning module is automatically triggered to update the classifier weights, forming a closed-loop optimization.

[0056] S7, when deployed on edge devices, the system starts dynamic computing path selection and evaluates the difficulty of the input image through a lightweight complexity prediction network. Simple samples (such as solid color backgrounds) skip 50% of the convolution layers and directly use shallow feature classification. Complex samples (such as dense objects) enable the full model. At the same time, FP16 precision is used to calculate the feature extraction front end, FP32 precision is retained in the classification layer, and asynchronous pipeline processing is used to achieve an average end-to-end delay of 45ms.

[0057] Furthermore, when the input data contains associated text information, the cross-modal joint analysis process is initiated. First, image features and text features are extracted respectively through independent visual encoders and text encoders. The visual encoder adopts a multi-scale feature extraction network, while the text encoder implements semantic embedding based on the pre-trained BERT model. Subsequently, a visual-semantic association model is constructed through a graph neural network, and image region features and text keywords are mapped to a unified graph structure. At the same time, the graph attention mechanism is used to learn the relationship weights between modalities. Finally, the visual features and text features are weightedly spliced through the cross-modal feature fusion layer and input into the multimodal classifier for joint prediction. The classifier uses bilinear pooling technology to capture high-order interaction features between modalities, and supports two working states: pure visual mode and visual-text hybrid mode. When text information is missing, it automatically degenerates to the standard image classification process.

[0058] Furthermore, the edge computing deployment solution first uses knowledge distillation technology to compress the original large classification model into a lightweight version, wherein the teacher model adopts a network of interpretable classification method, and the student model adopts a micro-architecture constructed by deep separable convolution, and realizes efficient knowledge transfer through feature map matching loss and classification logits distillation loss. The compressed model is then fixed-point quantized, and the 32-bit floating-point parameters are converted into 8-bit integer representation and cooperated with the dynamic range calibration algorithm to keep the accuracy loss within 1%. At runtime, the system dynamically adjusts the calculation path according to the resource status of the edge device, automatically enables fast inference mode for low-complexity images to skip some network layers, and integrates hardware acceleration interface, optimizes convolution operations for ARM architecture processors and supports GPU / NPU dedicated instruction set acceleration.

Claims

1. Image classification system based on image recognition technology, characterized by: Includes: Image acquisition module: used to obtain the original image data to be classified; Preprocessing module: denoising, normalizing and size standardizing the image; Feature extraction module: uses deep convolutional neural network to extract multi-level features of images; Classification decision module: weighted fusion features based on the attention mechanism and output classification results; Output module: displays classification labels and confidence scores.

2. The image classification system based on image recognition technology according to claim 1, characterized in that: The preprocessing module dynamically selects a denoising algorithm based on image quality assessment, selects a denoising strategy by calculating the signal-to-noise ratio of an image block, and simultaneously performs adaptive brightness correction and adjusts low-contrast images through histogram equalization.

3. The image classification system based on image recognition technology according to claim 1, characterized in that: The feature extraction module uses a multi-scale feature extraction network, which includes parallel convolution paths and feature pyramid structures. The parallel convolution paths are used to extract local features of different receptive fields. The first path is a 3×3 convolution kernel with a step size of 1 and outputs a feature map of size H×W×64. The second path is a 5×5 dilated convolution that outputs a feature map of the same size. The feature pyramid structure is used to fuse shallow texture features with deep semantic features.

4. The image classification system based on image recognition technology according to claim 1, characterized in that: The classification decision module uses an explainable classification method for heat map generation and multi-class classification output. The heat map generation is based on the Grad-CAM++ algorithm to calculate the category-sensitive area. The resolution of the heat map maintains a 1:4 ratio with the input image, and Jet color mapping is used for overlay display. The output of the multi-class classification is divided into three levels: the main classification result, the top-1 classification label and its confidence score. For secondary candidate results, when the Top-1 confidence is less than 0.7, the Top-3 categories and their probability distribution are output; for uncertainty annotation, samples with confidence less than 0.5 are marked as "low reliability".

5. The image classification system based on image recognition technology according to claim 1, characterized in that: The system also includes a new category learning unit and a sample memory library. The new category learning unit realizes online model update by freezing the underlying network layer and fine-tuning the classifier; the sample memory library includes a storage strategy, dynamic update and playback mechanism; Storage strategy: Select representative samples based on feature space K-means clustering; Dynamic update: When the distance between a new sample and an existing cluster center is greater than the threshold, it is stored in the memory bank; Replay mechanism: During training, new data and memory samples are mixed in a 1:1 ratio.

6. An image classification method based on image recognition technology, using the image classification system based on image recognition technology according to claims 1-5, characterized in that: The following steps are included: S1, first obtains the raw image data to be classified through the image acquisition module, and then enters the preprocessing module for automated processing. This module first evaluates the image quality to determine the appropriate denoising algorithm. For low-quality images, non-local mean filtering is used, while for high-quality images, fast bilateral filtering is used. At the same time, adaptive brightness correction is performed through CLAHE algorithm partition processing and combined with nonlinear brightness compression function to adjust the overexposed area. Finally, geometric correction is performed to eliminate image tilt and perspective distortion to ensure that the input image meets the standard requirements of subsequent feature extraction; S2, the preprocessed image is input into the hybrid neural network architecture of the feature extraction module, which consists of a combination of a ResNet-50 backbone network and an inserted Transformer layer. First, a parallel convolution path is used to extract local features of different receptive fields using 3×3 standard convolution and 5×5 dilated convolution respectively. Then, shallow texture features and deep semantic features are fused from top to bottom in the feature pyramid structure, and finally a multi-scale feature map with four levels from P2 to P5 is output, where high-level features contain rich semantic information while low-level features retain fine spatial details. S3, the dual attention mechanism that inputs the extracted multi-scale features into the classification decision module, first recalibrates the importance of the channel dimension of the feature map through the channel attention mechanism to suppress the response of irrelevant channels, then uses the spatial attention mechanism to focus on the key areas related to classification in the spatial dimension, and generates a spatial attention weight map through cascaded global average pooling and maximum pooling operations. Finally, the weighted feature map is input into the subsequent classifier; In S4, the attention-weighted features are first compressed in spatial dimensions through a global average pooling layer, and then input into a classifier containing a fully connected layer for category prediction. The output is calibrated using a temperature-adjusted Softmax function to balance the confidence distribution of each category. At the same time, the cross-entropy loss is calculated for parameter optimization in the model training phase. In the inference phase, the top-1 classification label and its confidence score are output. When the confidence of the main classification result is lower than the threshold, the top-3 candidate categories and their probability distribution are provided, and a visual heat map is generated through the Grad-CAM++ algorithm to enhance the interpretability of the classification results.

7. The image classification method based on image recognition technology according to claim 6, characterized in that: After the S1 step is completed, the input image is fast Fourier transformed and the high-frequency component features are extracted. When abnormal high-frequency energy distribution is detected, the adversarial sample alarm is triggered, and then the pixel-level defense process is started. This process uses an image reconstructor based on a generative adversarial network to repair local areas of suspected adversarial samples. The reconstructor achieves semantic-level reconstruction of damaged pixels through a pre-trained U-Net architecture. At the same time, randomization processing is introduced in the feature extraction stage, including adding Gaussian noise perturbation to the weights of the convolutional layer and performing random spatial transformation on the feature map. Finally, a confidence threshold detection mechanism is deployed in the classification decision module. When the confidence of the output prediction is lower than the preset threshold, the sample is automatically marked as a suspicious adversarial sample and a secondary verification process is started. The robustness of the final classification result is ensured through a multi-model voting mechanism.

8. The image classification method based on image recognition technology according to claim 6, characterized in that: When the input data contains associated text information, the cross-modal joint analysis process is initiated. First, image features and text features are extracted respectively through independent visual encoders and text encoders. The visual encoder adopts a multi-scale feature extraction network, while the text encoder implements semantic embedding based on the pre-trained BERT model. Subsequently, a visual-semantic association model is constructed through a graph neural network, and image region features and text keywords are mapped to a unified graph structure. At the same time, the graph attention mechanism is used to learn the relationship weights between modalities. Finally, the visual features and text features are weightedly spliced through the cross-modal feature fusion layer and input into the multimodal classifier for joint prediction. The classifier uses bilinear pooling technology to capture high-order interaction features between modalities, and supports two working states: pure visual mode and visual-text hybrid mode. When text information is missing, it automatically degenerates to the standard image classification process.

9. The image classification method based on image recognition technology according to claim 6, characterized in that: The method also uses an edge computing deployment solution. First, the knowledge distillation technology is used to compress the original large classification model into a lightweight version. The teacher model adopts a network of an explainable classification method, and the student model adopts a micro-architecture constructed with deep separable convolution. Efficient knowledge transfer is achieved through feature map matching loss and classification logits distillation loss. The compressed model is then fixed-point quantized, and the 32-bit floating-point parameters are converted into 8-bit integer representation. The dynamic range calibration algorithm is used to keep the accuracy loss within 1%. At runtime, the system dynamically adjusts the calculation path according to the resource status of the edge device, automatically enables fast inference mode for low-complexity images, skips some network layers, and integrates a hardware acceleration interface to optimize convolution operations for ARM architecture processors and support dedicated instruction set acceleration for GPU / NPU.

10. The image classification method based on image recognition technology according to claim 6, characterized in that: Dynamic computation paths are enabled during the inference phase, skipping redundant network layers for simple samples while using mixed-precision computation to accelerate feature extraction.

Images