Intelligent image recognition and classification system based on deep learning

Abstract

The invention relates to the technical field of computer vision and artificial intelligence, and discloses an intelligent image recognition and classification system based on deep learning, which comprises an image preprocessing module, a feature extraction module, a semantic modeling module, a classification decision module and a self-adaptive optimization module. The system firstly performs standardization processing on an original image, extracts multi-scale texture and edge features, and models context semantic information in combination with an attention mechanism to form high-dimensional representation; and then multi-label output is realized through nonlinear classification mapping, and parameter weights are adaptively optimized according to prediction errors. The system has high robustness and generalization ability in a multi-category image scene, and is suitable for intelligent identification and refined classification tasks in a complex image environment.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and in particular to an image recognition and classification system based on deep learning. Background Art

[0002] With the rapid development of deep learning technology, neural network-based image recognition and classification systems have been widely used in a variety of fields, including industrial inspection, security monitoring, and medical diagnosis. Although current research has made significant progress in image feature representation, semantic modeling, and decision-making mechanisms, existing technologies still have many shortcomings in complex real-world scenarios, such as inconsistent image resolution, insufficient feature extraction, weak context modeling, poor classification generalization, and unstable model training. They are unable to effectively support the needs of high-precision image recognition and multi-label classification in complex environments.

[0003] Existing image recognition systems typically perform fixed-scale normalization, grayscale conversion, or histogram equalization on input images, failing to adaptively enhance or adjust resolution for different image scenarios. For example, some images with uneven lighting or local blur may lose important feature areas, affecting subsequent recognition accuracy. Furthermore, the resolution of images from different sources varies significantly, and traditional unified scaling schemes cannot guarantee consistent feature representation. This can easily lead to model sensitivity to scale changes and a lack of robustness in front-end standardization.

[0004] Current image classification systems primarily rely on forward reasoning for semantic extraction and decision output, lacking an optimization loop based on error feedback. This results in classification performance being highly sensitive to the network's initial parameters. In multi-label classification, in particular, models may overfit to some labels and lack the ability to recognize less frequent labels. Furthermore, some systems fail to fully leverage contextual relationships to model semantic connections, resulting in a tendency for isolated semantic feature representations. Furthermore, they lack cross-layer attention mechanisms and temporal integration capabilities, making robust image understanding and accurate classification difficult. Summary of the Invention

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solution: an intelligent image recognition and classification system based on deep learning, comprising the following modules:

[0006] Image preprocessing module, used to receive original image data I raw , and normalize, enhance and adjust the resolution to output standardized image data I std ;

[0007] Feature extraction module, used to receive standardized image data I std , extract local texture features F based on multi-scale convolution kernel tex and edge features F edg , forming a fusion feature vector Fext ;

[0008] a semantic modeling module configured to receive the fusion feature vector F ext , extract a context-related semantic tensor S ctx by a dual-channel attention network, and generate an intermediate representation encoding S emb ;

[0009] a classification decision module configured to receive the intermediate representation encoding S emb , calculate a probability distribution P cls using a nonlinear mapping function, realize multi-class label output, and calculate an error term ε cls based on the probability distribution P gt and a true label P cls ;

[0010] an adaptive optimization module configured to update each learnable parameter weight in the system based on the error term ε cls .

[0011] Preferably, the image preprocessing module comprises:

[0012] an image enhancement unit configured to perform an adaptive correction enhancement algorithm on the original image data I raw , and the enhancement expression is:

[0013]

[0014] wherein γ val is a gamma correction index, γ val = 1 + 0.25σ roi , and σ roi is a brightness variance of the image ROI region, and the calculation expression is:

[0015]

[0016] L i is a brightness value of the i-th pixel, L roi is an average value of the brightness in the image ROI region, and N is a pixel number in the image ROI;

[0017] a resolution resampling unit configured to use a bilinear interpolation function to normalize the image to a unified input size H std × W std , so as to facilitate subsequent operation unified processing.

[0018] Preferably, the feature extraction module comprises:

[0019] a multi-scale convolution unit configured to use a set of convolution kernels with variable scales to extract local texture features F from the normalized image data I std .tex , and its calculation formula is:

[0020]

[0021] in, is the standard ReLU activation function, * is the convolution operation;

[0022] Edge enhancement unit, using Laplace edge operator K lap Perform edge feature F edg Extract, the Laplace edge operator and local texture features F tex Splicing to form a fusion feature vector F ext :

[0023] F ext =Concat(F tex ,F edg );

[0024] Among them, F edg =K lap *I std ;

[0025] Preferably, the semantic modeling module includes:

[0026] Channel attention modeling unit, using SE structure to fusion feature vector F ext Channel weighting generates channel attention vector ω ch :

[0027]

[0028] Among them, σ is the Sigmoid activation function, is the standard ReLU activation function, ω1 and ω2 are the learnable weight parameters of the current channel, and satisfy ω1+ω2=1, F ext is the fused feature vector, H is the height of the image, and W is the width of the image.

[0029] The spatial context modeling unit adopts a position-sensitive attention mechanism to convert the channel attention vector ω ch Acting on the fusion feature vector F ext , construct the context semantic tensor S ctx , and use bidirectional gated recurrent function to perform semantic temporal modeling to generate intermediate representation encoding S emb :

[0030] S emb =Bi-GRU(S ctx );

[0031] Among them, Bi-GRU is a standard bidirectional gated recurrent function.

[0032] Preferably, the classification decision module includes:

[0033] Mapping function building unit, defining nonlinear classification mapping as probability distribution P cls :

[0034]

[0035] Among them, ω j is the weight parameter of the jth class, and satisfies C is the total number of classification categories, S emb Encode the intermediate representation;

[0036] The multi-label decoding unit outputs the classification result based on the label with the highest probability, and provides supervision error information for the next adaptive optimization module:

[0037]

[0038] in, is the true label value of the kth category, ranging from 0 to 1, is the model's predicted probability for the kth class, and K represents the total number of labels.

[0039] Preferably, the adaptive optimization module includes:

[0040] Parameter updating unit, used to update the error term ε output by the classification decision module cls , for the learnable weight parameter ω in the model t Perform iterative nonlinear optimization updates, combined with adaptive learning rate η t The update rule with momentum factor μ is:

[0041]

[0042] Among them, ω t is the learning weight in the corresponding model, ω t+1 is the learning weight in the updated model

[0043] represents the error term ε cls Weight ω t The first-order partial derivative of the momentum factor μ is in the range [0,1], η t The learning rate is dynamically adjusted according to the root mean square error gradient and is defined as:

[0044]

[0045] η0 is the initial learning rate, the default value is 0.2;

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] Improve the adaptability of image input and enhance the robustness of front-end preprocessing: This invention constructs an image preprocessing module and adopts multi-scale adaptive filtering and illumination enhancement algorithms based on image distribution characteristics. It significantly enhances image quality while retaining the original structural information, effectively solving the problem of feature loss caused by inconsistent image resolution, local blur, and uneven illumination in the existing technology, and improves the system's compatibility and recognition stability for image input in complex environments.

[0048] Enhance feature extraction and semantic understanding capabilities, and achieve multi-level deep semantic fusion: This invention integrates an improved residual attention network with a multi-scale feature encoding mechanism to achieve joint modeling of texture, edge, and contextual semantics. While maintaining the accuracy of low-level structures, it introduces cross-layer semantic guidance, solving the problems of lack of detailed support for high-level feature expression and isolated semantic structures in existing image recognition systems, and significantly improving the model's ability to discriminate complex images.

[0049] Introducing an adaptive feedback optimization mechanism to improve classification accuracy and system generalization ability: The present invention introduces a feedback mechanism based on dynamic loss adjustment and sample gradient weighting in the classification decision module and the adaptive optimization module, constructs a model iteration path driven by classification errors, and constructs an optimized graph structure based on the correlation between labels. This breaks through the limitation of traditional static classifiers that cannot automatically correct errors, and effectively improves the classification accuracy and generalization ability of the model in multi-label and unbalanced sample scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of the system module flow provided for this application;

[0051] Figure 2 Schematic diagram of the image preprocessing module provided in this application;

[0052] Figure 3 Schematic diagram of the feature extraction module provided by this application;

[0053] Figure 4 Schematic diagram of the semantic modeling module provided for this application;

[0054] Figure 5 Schematic diagram of the classification decision module provided for this application;

[0055] Figure 6 Schematic diagram of the adaptive optimization module provided in this application. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the solutions of the present invention, 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 embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0057] refer to Figures 1-6 , an embodiment of the present invention provides an intelligent image recognition and classification system based on deep learning, including the following modules:

[0058] Image preprocessing module, used to receive original image data I raw , and normalize, enhance and adjust the resolution to output standardized image data I std .

[0059] The image preprocessing module mainly consists of an image enhancement unit and a resolution resampling unit, which are used to perform uniform standardization and quality enhancement on the original image data to ensure the robustness of subsequent feature extraction and the uniformity of model input. Specifically, the system first obtains the original image data I from the image acquisition terminal. raw In natural scene acquisition, the image often has problems such as local underexposure, uneven illumination or color deviation. The image enhancement unit uses an adaptive gamma correction method based on brightness variance to dynamically enhance the image. First, the brightness value sequence L is extracted in the ROI area. i , calculate its mean L roi and variance Then dynamically adjust the gamma index γ val =1+0.25σ roi , and finally according to the enhanced expression The original image is enhanced pixel by pixel, making the overall contrast and texture details clearer while retaining the structural features. After the enhancement process, the resolution resampling unit uses a bilinear interpolation algorithm to standardize the image to the preset input size H. std ×W std (e.g., 224×224), ensuring that the data structure of each sample is consistent when the model is batch-input, effectively improving the stability and computational efficiency of subsequent network processing. Through this adaptive image preprocessing mechanism, the system significantly improves its adaptability and data quality for image input in complex environments, ensuring the effectiveness of the deep feature extraction stage.

[0060] Feature extraction module, used to receive standardized image data I std , extract local texture features F based on multi-scale convolution kernel tex and edge features Fedg , forming a fusion feature vector F ext .

[0061] The feature extraction module consists of a multi-scale convolution unit and an edge enhancement unit, which is designed to extract the feature from the pre-processed standardized image data I std Multi-level texture information extraction and structural edge enhancement are performed to improve the discrimination ability and robustness of the subsequent classification model for the target image. First, the multi-scale convolution unit introduces a set of convolution kernels with variable scales. The convolution kernel sizes range from small (such as 3×3, 5×5) to large (such as 7×7, 11×11), aiming to adapt to the texture changes at different scales in the image. Each convolution kernel K si Will act independently on the input image I std , perform standard convolution operation K si *I std , then the nonlinear feature response is retained by the ReLU activation function to form local response maps at different scales. The activation response results at all scales are finally fused into a unified local texture feature map by channel-by-channel summation. This feature map integrates multiple scale information while maintaining the consistency of spatial position, and can effectively characterize local texture differences, repeated structures and weak pattern changes in the image. Subsequently, the image is passed to the edge enhancement unit to further enhance its structural contour features. This unit is based on the standard Laplacian edge operator. For image I std Perform second-order derivative convolution operation F edg =K lap *I std , extract the edge contour information in the image. Since the Laplace operator is sensitive to high-frequency components (such as edges and mutation areas), the generated F edg The feature map can highlight the significant edges, structural junctions and contour change areas in the image, which supplements the limitations of the convolution operation in texture modeling. In the feature fusion stage, the system uses the channel dimension splicing operation to combine the two feature maps F tex and F edg Merge into a unified fusion feature vector F ext =Concat(F tex ,F edg ), forming a high-dimensional feature description that incorporates multi-scale texture information while also taking into account edge structure. This fused feature, when subsequently fed into the classification and discrimination module, not only enhances the model's overall perception of image spatial distribution, texture complexity, and shape contours, but also improves the model's robustness and accuracy in complex scenes such as blurred backgrounds, overlapping structures, and changing lighting.

[0062] Semantic modeling module, used to receive the fused feature vector Fext , extracting context-dependent semantic tensor S through a dual-channel attention network ctx , and generate the intermediate representation code S emb .

[0063] The semantic modeling module consists of a channel attention modeling unit and a spatial context modeling unit. Its main task is to ext Semantic enhancement and sequence encoding are performed to extract high-level discriminative information and retain the semantic relevance of spatial structure. First, the channel attention modeling unit models the importance of feature channels by introducing the SE structure. Specifically, the system first performs the fusion feature vector F ext Perform global average pooling in the spatial dimension, that is, average the H×W pixels of each channel to form a channel description vector z. The core calculation formula of this process is:

[0064]

[0065] Subsequently, the system introduces two fully connected layer parameters ω1 and ω2, where ω1+ω2=1, to learn the nonlinear relationship between channels. First, z is multiplied by the weight parameter ω1 and activated by the standard ReLU function. Extract the intermediate expression, input the result into the second layer weight ω2, and finally generate the channel attention vector ω through the Sigmoid activation function σ ch , which is used to reflect the importance of each channel in the overall feature expression. Its mathematical expression is:

[0066]

[0067] Generated channel attention vector ω ch With F ext The same channel dimension, by ext After performing channel-by-channel multiplication, the significant semantic channels can be strengthened, thereby forming a semantic mapping tensor with channel weighting characteristics. Next, it enters the spatial context modeling unit, which uses the position-sensitive attention mechanism to perform spatial modeling on the fused feature map. First, the channel attention vector ω obtained in the previous step is ch Re-act on the fused feature vector F ext , that is, the feature responses at different positions are adjusted by channel weighting to construct the spatial context semantic tensor S ctx This tensor not only preserves the structural relationship of local spatial positions, but also introduces the semantic enhancement information after channel weighting, thus having richer context expression capabilities.

[0068] In order to further extract the potential sequence dependency of images in the spatial dimension, the system will ctxThe input is fed into a standard bidirectional gated recurrent unit (Bi-GRU) for sequence modeling. Bi-GRU consists of two GRU units in opposite directions, traversing the spatial sequence from left to right and right to left respectively, and can capture the global semantic association between the previous and next states. The intermediate representation encoding of its output is denoted by S emb =Bi-GRU(S ctx ), which contains the sequential semantics, structural features and contextual information of the image in the spatial dimension, which helps to achieve more accurate feature discrimination and target recognition in the subsequent classification module.

[0069] Classification decision module, used to receive the intermediate representation code S emb , using nonlinear mapping function to calculate the probability distribution P cls , realize multi-category label output, and according to the probability distribution P cls and the true label P gt Calculate the error term ε cls .

[0070] The classification decision module is mainly composed of a mapping function construction unit and a multi-label decoding unit, which aims to transform the deep feature vector S emb It is converted into a clear classification result and outputs the supervised error information for subsequent optimization. First, the mapping function construction unit converts the intermediate representation encoding S into emb Converted into classification probability distribution P cls The mapping function adopts the softmax structure, which is weighted by each category ω j With S emb The linear relationship between them is exponentially mapped and normalized to generate the predicted probability of the corresponding category. The specific mapping function formula is:

[0071]

[0072] Among them, ω j is the learnable weight parameter of the jth class, and all class weights satisfy the normalization constraint C is the total number of categories set in the image classification task, for example, 10 types of traffic signs, 20 types of industrial defect types, or 1000 types of object recognition labels. Through the above method, the system can effectively map high-dimensional embedded features into multi-class probability outputs, improving the discriminability and interpretability in multi-classification tasks. Subsequently, the multi-label decoding unit in the classification decision module generates a probability distribution P. clsDetermine the final classification result. Specifically, this unit first sorts the probability values ​​of each predicted category and takes the category corresponding to the maximum probability as the identification label output of the current image. In addition, in order to achieve end-to-end optimization training of the model, the multi-label decoding unit is also responsible for calculating the supervision error information, which is specifically described by the cross entropy loss function. Its formula is as follows:

[0073]

[0074] in, represents the true label of the kth category, and its value range is [0,1]; is the predicted probability value of the system's current output for the kth class, where K represents the total number of labels, which is usually equivalent to the total number of classification categories C. This loss function reflects the degree of deviation between the current output of the model and the true label, serving as a key error signal for subsequent backpropagation training.

[0075] In actual deployment, the classification decision module can support different types of image recognition tasks, such as identity classification in face recognition, defect location classification in industrial visual inspection, or lesion classification in medical images. With the help of the softmax normalization strategy of the mapping function construction unit and the cross entropy error feedback mechanism of the multi-label decoding unit, the system has strong discrimination ability and scalability. At the same time, the weight normalization constraint adopted by the module It also ensures that the model maintains a balanced distribution among categories during the classification process, avoiding the bias problem caused by category imbalance.

[0076] Adaptive optimization module for cls Update the weights of each learnable parameter in the system.

[0077] In the adaptive optimization module, the parameter update unit is based on the supervised error term ε output by the classification decision module. cls , combined with the adaptive learning rate η t With the momentum factor μ, an iterative update rule with nonlinear characteristics is implemented. First, for the current learnable parameter ω in each training round t , the system will perform a first-order derivative of the gradient of the parameter under the current error background to obtain the error term ε cls Weight ω t The first-order partial derivative of Subsequently, the momentum factor μ∈[0,1] is introduced to fuse the weight update direction of the previous round with the current gradient direction, aiming to use historical information to smooth the update path, thereby avoiding parameter oscillation in the local minimum area and ensuring the continuity and stability of the optimization process.

[0078] In terms of learning rate, this embodiment does not adopt a fixed learning rate strategy, but introduces an adaptive learning rate η t , the learning rate is dynamically adjusted according to the current gradient norm and is defined as:

[0079]

[0080] Here, η0 is the initial learning rate, which is set to 0.2 by default in this system. This formula effectively suppresses gradient oscillations by automatically reducing the step size when the gradient is large to prevent overly aggressive updates; and increasing the step size when the gradient is small to accelerate convergence, thereby improving overall training efficiency.

[0081] The final parameter update formula is as follows:

[0082]

[0083] Here, through the momentum term μ·ω t With dynamic learning items The weighted combination of η makes the model weight update path maintain the consistency of the historical optimization direction and have the dynamic response capability that is sensitive to the current error. When the error in the initial stage of the model is large, η t Small, the update step size is controlled, effectively avoiding overfitting; as the error gradually converges, the gradient term decreases, η t Increasing it speeds up the convergence.

[0084] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.

[0085] Obviously, the embodiments described above are only some embodiments of the present invention, rather than all embodiments. The preferred embodiments of the present invention are given in the accompanying drawings, but they do not limit the patent scope of the present invention. The present invention can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive. Although the present invention has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present invention specification and drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present invention.

Claims

1. An intelligent image recognition and classification system based on deep learning, characterized in that: Includes the following modules: Image preprocessing module, used to receive original image data I raw , and normalize, enhance and adjust the resolution to output standardized image data I std ; Feature extraction module, used to receive standardized image data I std , extract local texture features F based on multi-scale convolution kernel tex and edge features F edg , forming a fusion feature vector F ext ; Semantic modeling module, used to receive the fused feature vector F ext , extracting context-dependent semantic tensor S through a dual-channel attention network ctx , and generate the intermediate representation code S emb ; Classification decision module, used to receive the intermediate representation code S emb , using nonlinear mapping function to calculate the probability distribution P cls , realize multi-category label output, and according to the probability distribution P cls and the true label P gt Calculate the error term ε cls ; Adaptive optimization module for cls Calculate the model gradient and combine the adaptive learning rate and momentum coefficient to perform nonlinear update optimization on the model parameters.

2. The deep learning-based intelligent image recognition and classification system according to claim 1, characterized in that: The image preprocessing module includes: Image enhancement unit, used to enhance the original image data I raw Execute the adaptive correction enhancement algorithm, and the enhancement expression formula is: Among them, γ val is the gamma correction index, γ val =1+0.25σ roi , σ roi is the brightness variance of the image ROI area, and the calculation formula is: L i is the brightness value of the i-th pixel, L roi is the average brightness within the image ROI area, and N is the number of pixels in the image ROI; Resolution resampling unit, which uses a bilinear interpolation function to normalize the image to a uniform input size H std ×W std , which facilitates unified processing of subsequent operations.

3. The deep learning-based intelligent image recognition and classification system according to claim 1, characterized in that: The feature extraction module includes: Multi-scale convolution unit, using a set of convolution kernels with variable scale For the standardized image data I std Extract local texture features F tex , and its calculation formula is: in, is the standard ReLU activation function, * is the convolution operation; Edge enhancement unit, using Laplace edge operator K lap Perform edge feature F edg Extract, the Laplace edge operator and local texture features F tex Splicing to form a fusion feature vector F ext : F ext =Concat(F tex ,F edg ); Among them, F edg =K lap *I std。 4. The deep learning-based intelligent image recognition and classification system according to claim 1, characterized in that: The semantic modeling module includes: Channel attention modeling unit, using SE structure to fusion feature vector F ext Channel weighting generates channel attention vector ω ch : Among them, σ is the Sigmoid activation function, is the standard ReLU activation function, ω1 and ω2 are the learnable weight parameters of the current channel, and satisfy ω1+ω2=1, F ext is the fused feature vector, H is the height of the image, and W is the width of the image. The spatial context modeling unit adopts a position-sensitive attention mechanism to convert the channel attention vector ω ch Acting on the fusion feature vector F ext , construct the context semantic tensor S ctx , and use bidirectional gated recurrent function to perform semantic temporal modeling to generate intermediate representation encoding S emb : WITH emb =Bi-GRU(S ctx ); Among them, Bi-GRU is a standard bidirectional gated recurrent function.

5. The deep learning-based intelligent image recognition and classification system according to claim 1, characterized in that: The classification decision module includes: Mapping function building unit, defining nonlinear classification mapping as probability distribution P cls : Among them, ω j is the weight parameter of the jth class, and satisfies C is the total number of categories, S emb Encode the intermediate representation; The multi-label decoding unit outputs the classification result based on the label with the highest probability, and provides supervision error information for the next adaptive optimization module: in, is the true label value of the kth category, ranging from 0 to 1, is the model's predicted probability for the kth class, and K represents the total number of labels.

6. The deep learning-based intelligent image recognition and classification system according to claim 1, characterized in that: The adaptive optimization module includes: Parameter updating unit, used to update the error term ε output by the classification decision module cls , for the learnable weight parameter ω in the model t Perform iterative nonlinear optimization updates, combined with adaptive learning rate η t The update rule with momentum factor μ is: Among them, ω t is the learning weight in the corresponding model, ω t+1 is the learning weight in the updated model represents the error term ε cls Weight ω t The first-order partial derivative of the momentum factor μ is in the range [0,1], η t The learning rate is dynamically adjusted according to the root mean square error gradient and is defined as: η0 is the initial learning rate, the default value is 0.2.

Images