An intelligent image recognition and classification system based on deep learning

Abstract

The application 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, combines an attention mechanism to model context semantic information, and forms high-dimensional representation; then multi-label output is realized through nonlinear classification mapping, and parameter weights are self-adaptively optimized according to a prediction error. The system has high robustness and generalization ability in a multi-class image scene, and is suitable for intelligent recognition and fine classification tasks in a complex image environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and artificial intelligence, and particularly relates to an image recognition and classification system based on deep learning. BACKGROUND

[0002] With the rapid development of deep learning technology, image recognition and classification systems based on neural networks have been widely applied in industrial detection, security monitoring, medical diagnosis and other fields. Although current research has made many progress in image feature expression, semantic modeling and decision mechanism, in the face of inconsistent image resolution, insufficient feature extraction, weak context modeling ability, poor classification generalization ability and unstable model training, the existing technology still has many deficiencies, which cannot effectively support the demand of high-precision image recognition and multi-label classification in complex environment.

[0003] The existing image recognition system usually performs fixed scale normalization, gray scale conversion or histogram equalization processing on the input image, and cannot adaptively enhance or adjust the resolution according to different image scenes. For example, some images will lose important feature regions when the light is uneven or locally blurred, which will affect the subsequent recognition accuracy. In addition, there are significant differences in the resolution of images from different sources, and the traditional uniform scaling scheme cannot guarantee consistent expression of features, which is easy to cause the model to be sensitive to scale changes and lack of front-end standardization robustness.

[0004] The current image classification system mainly relies on forward reasoning for semantic extraction and decision output, and lacks an optimization loop based on error feedback, which makes the classification performance highly sensitive to the initial parameters of the network. Especially in multi-label classification, the model may overfit to some labels, while the recognition ability for low-frequency labels is insufficient. In addition, some systems do not fully utilize the context relationship modeling semantic connection, and the semantic feature expression has a tendency to isolate, lacking cross-layer attention mechanism and time integration ability, making it difficult to achieve robust image understanding and accurate classification. SUMMARY

[0005] In order to achieve the above application purpose, the present application provides the following technical scheme: an intelligent image recognition and classification system based on deep learning, comprising the following modules:

[0006] An image preprocessing module is used to receive original image data I raw , and perform normalization, enhancement and resolution adjustment on it, and output standardized image data I std ;

[0007] A feature extraction module is used to receive standardized image data I std , extract local texture features F tex and edge features F edg based on multi-scale convolution kernels, and form a fusion feature vector Fext The feature extraction module includes:

[0008] Multi-scale convolutional units employ a set of scale-variable convolutional kernels. For standardized image data I std Extracting local texture features F tex The calculation formula is as follows:

[0009]

[0010] in, The standard ReLU activation function is used, and * represents the convolution operation;

[0011] Edge enhancement unit, using Laplacian edge operator K lap Perform edge feature F edg Extraction, the Laplacian edge operator and with local texture features F tex The fused feature vector F is formed by splicing the features together. ext :

[0012] F ext =Concat(F tex ,F edg );

[0013] Among them, F edg =K lap *I std ;

[0014] The semantic modeling module is used to receive the fused feature vector F. ext The context-related semantic tensor S is extracted through a dual-channel attention network. ctx And generate intermediate representation encoding S emb ;

[0015] The classification decision module is used to receive the intermediate representation code S. emb The probability distribution P is calculated using a nonlinear mapping function. cls To achieve multi-category label output, and based on probability distribution P cls and real label P gt Calculate the error term ε cls ;

[0016] An adaptive optimization module is used to optimize based on the error term ε. cls The adaptive optimization module updates the weights of each learnable parameter in the system, and includes:

[0017] The parameter update unit is used to update the error term ε output by the classification decision module. cls For the learnable weight parameters ω in the model titerative nonlinear optimization update combined with an adaptive learning rate η t The update rule of the momentum factor μ is:

[0018]

[0019] where ω t is the learning weight in the corresponding model, ω t+1 is the learning weight in the updated model

[0020] represents the error term ε cls , the first-order partial derivative of the weight ω t , the momentum factor μ is in the range [0, 1], η t is a learning rate dynamically adjusted according to the root mean square of the error gradient, defined as:

[0021]

[0022] η0 is the initial learning rate, and the default value is 0.2;

[0023] Preferably, the image preprocessing module comprises:

[0024] An image enhancement unit is configured to perform an adaptive correction enhancement algorithm on the original image data I raw , and the enhancement expression formula is:

[0025]

[0026] where γ val is a gamma correction index, γ val = 1 + 0.25σ roi , σ roi is the brightness variance of the image ROI region, and the calculation formula is:

[0027]

[0028] L i is the brightness value of the i-th pixel, L roi is the average brightness value in the image ROI region, and N is the number of pixels in the image ROI.

[0029] A resolution resampling unit is configured to use a bilinear interpolation function to normalize the image to a unified input size H std × W std for subsequent operation and unified processing.

[0030] Preferably, the semantic modeling module comprises:

[0031] A channel attention modeling unit is configured to use an SE structure to process the fused feature vector F extChannel weighted generation of channel attention vector ω ch :

[0032]

[0033] Where σ is the Sigmoid activation function. The function is a standard ReLU activation function, where ω1 and ω2 are the learnable weights for the current channel, and satisfy ω1 + ω2 = 1. F ext To fuse feature vectors, H is the height of the image and W is the width of the image.

[0034] The spatial context modeling unit employs a position-sensitive attention mechanism, which integrates the channel attention vector ω. ch Acting on the fused feature vector F ext Construct the context semantic tensor S ctx And a bidirectional gated loop function is used for semantic temporal modeling to generate intermediate representation encoding S. emb :

[0035] S emb =Bi-GRU(S) ctx );

[0036] Bi-GRU is a standard bidirectional gated loop function.

[0037] Preferably, the classification decision module includes:

[0038] The mapping function building block defines the nonlinear classification mapping as a probability distribution P. cls :

[0039]

[0040] Where, ω j Let be the weight parameters of the j-th class, and satisfy . C represents the total number of categories, S emb Encoding the intermediate representation;

[0041] 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.

[0042]

[0043] in, The true label value for the k-th class, ranging from 0 to 1. Let K be the predicted probability of the model for the k-th class, where K represents the total number of labels.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] Improve image input adaptability, enhance front-end preprocessing robustness: the present application constructs an image preprocessing module, adopts multi-scale adaptive filtering and illumination enhancement algorithm based on image distribution characteristics, significantly enhances image quality while preserving original structure information, effectively solves the feature missing problem caused by inconsistent image resolution, local blur and uneven illumination in the prior art, and improves the compatibility and recognition stability of the system for image input in complex environment.

[0046] Enhance feature extraction and semantic understanding ability, realize multi-level deep semantic fusion: the present application realizes joint modeling of texture, edge and context semantics by fusing improved residual attention network and multi-scale feature encoding mechanism, introduces cross-layer semantic guidance while maintaining low-layer structure precision, solves the problems of lack of detail support for high-level feature expression and isolated semantic structure in existing image recognition systems, and significantly improves the discrimination ability of the model for complex images.

[0047] Introduce adaptive feedback optimization mechanism, improve classification accuracy and system generalization ability: the present application introduces a feedback mechanism based on dynamic loss adjustment and sample gradient weighting in the classification decision module and adaptive optimization module, constructs a model iteration path driven by classification errors, and constructs an optimization graph structure combined with label correlation, breaks through the limitation that traditional static classifiers cannot automatically correct errors, and effectively improves the classification accuracy and generalization ability of the model in the multi-label and unbalanced sample scene. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The system module flowchart provided by the present application is provided;

[0049] Figure 2 The image preprocessing module provided by the present application is provided;

[0050] Figure 3 The feature extraction module provided by the present application is provided;

[0051] Figure 4 The semantic modeling module provided by the present application is provided;

[0052] Figure 5 The classification decision module provided by the present application is provided;

[0053] Figure 6 The adaptive optimization module provided by the present application is provided. DETAILED DESCRIPTION

[0054] In order to make the person skilled in the art better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the present application.

[0055] Reference Figures 1-6 The embodiments of the present application provide an intelligent image recognition and classification system based on deep learning, comprising the following modules:

[0056] An image preprocessing module is used to receive original image data I raw , and normalize, enhance and adjust the resolution of the original image data, and output standardized image data I std .

[0057] The image preprocessing module mainly consists of an image enhancement unit and a resolution resampling unit, which are used to uniformly standardize and quality enhance the original image data, to ensure the robustness of subsequent feature extraction and the uniformity of model input. Specifically, the system first acquires original image data I raw from an image acquisition terminal. The image often has problems such as local underexposure, uneven lighting or color deviation in natural acquisition scenarios. The image enhancement unit uses an adaptive gamma correction method based on luminance variance driving to dynamically enhance the image. First, the luminance value sequence L i is extracted in the ROI region, and the mean L roi and variance are calculated. Then the gamma index γ val is dynamically adjusted as γ roi =1+0.25σ roi , and finally the original image is enhanced pixel by pixel according to the enhancement expression , so that the image retains structural features while the overall contrast and texture details are clearer. After enhancement processing, the resolution resampling unit uses a bilinear interpolation algorithm to uniformly standardize the image to a preset input size H std ×W std (e.g. 224×224), to ensure that the model has consistent sample data structure when batch input, effectively improving the stability and computational efficiency of subsequent network processing. Through the above adaptive image preprocessing mechanism, the system significantly improves the adaptability and data quality of image input in complex environments, ensuring the effectiveness of the deep feature extraction stage.

[0058] A feature extraction module is used to receive standardized image data I std , extract local texture features F tex and edge features Fedg , forming a fusion feature vector F ext .

[0059] The feature extraction module is composed of a multi-scale convolution unit and an edge enhancement unit, aiming to extract multi-level texture information and strengthen structural edges of the pre-processed standardized image data I std to improve the discriminative ability and robustness of the subsequent classification model for the target image. First, the multi-scale convolution unit introduces a set of scale-variable convolution kernels K with sizes ranging from small (e.g. 3x3, 5x5) to large (e.g. 7x7, 11x11), aiming to adapt to texture changes at different scales in the image. Each convolution kernel K si acts independently on the input image I std , performing a standard convolution operation K si *I std , followed by a ReLU activation function to retain nonlinear feature responses, forming local response maps at different scales. The activation response results at all scales are finally fused into a unified local texture feature map F by channel-wise summation. This feature map, while maintaining consistent spatial positions, fuses multiple scale information and can effectively represent local texture differences, repeated structures, and subtle pattern changes in the image. Subsequently, the image is passed into the edge enhancement unit to further strengthen its structural contour features. Based on the standard Laplacian edge operator , the image I std is subjected to a second-order derivative convolution operation F edg = K lap *I std to extract edge contour information in the image. Since the Laplacian operator is sensitive to high-frequency components (such as edges and abrupt changes), the generated F edg feature map can highlight significant edges, structural boundaries, and contour change regions in the image, complementing the limitations of convolution operations in texture modeling. In the feature fusion stage, the system uses channel dimension concatenation to merge the two feature maps F tex and F edg into a unified fusion feature vector F ext = Concat(F tex , F edg ), forming a high-dimensional feature description result that contains both multi-scale texture information and edge structure. When this fusion feature is input into the subsequent classification and discrimination module, it not only enhances the model's overall perception of image spatial distribution, texture complexity, and shape contour, but also improves the model's discriminative robustness and accuracy in complex scenarios such as blurred background, structural overlap, or light changes.

[0060] The semantic modeling module is configured to receive the fusion feature vector Fext , extracts context-related semantic tensor S by a dual-channel attention network cxt , and generates intermediate representation code S emb .

[0061] The semantic modeling module is composed of a channel attention modeling unit and a spatial context modeling unit. The main task is to perform semantic enhancement and sequence coding on the aforementioned fusion feature vector F ext , to extract high-level discriminative information and preserve the semantic relevance of spatial structure. First, the channel attention modeling unit models the importance of feature channels by introducing an SE structure. Specifically, the system first performs global average pooling on the fusion feature vector F ext in the spatial dimension, i.e., averaging the HxW pixel points of each channel to form a channel description vector z. The core calculation formula of this process is:

[0062]

[0063] Subsequently, the system introduces two fully connected layer parameters ω1 and ω2, where ω1+ω2=1, to learn the non-linear relationship between channels. First, z is multiplied by the weight parameter ω1, and the standard ReLU activation function is used to extract the intermediate expression, and then the result is input to the second layer weight ω2, and finally the Sigmoid activation function σ is used to generate the channel attention vector ω ch , which reflects the importance of each channel in the overall feature expression. Its mathematical expression is:

[0064]

[0065] The generated channel attention vector ω ch has the same channel dimension as F ext . After performing channel-by-channel multiplication with F ext , 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 a position-sensitive attention mechanism to model the spatial features of the fused feature map. First, the channel attention vector ω ch obtained in the previous step is re-applied to the fusion feature vector F ext , i.e., adjusting the feature responses at different positions through channel weighting to construct a spatial context semantic tensor S ctx . This tensor not only preserves the structural relationship of local spatial positions, but also introduces semantic enhancement information after channel weighting, thereby having richer context expression capability.

[0066] To further extract the potential sequence-dependent relationship of the image in the spatial dimension, the system will S ctxThe input is fed into a standard bidirectional gated recurrent unit (Bi-GRU) for sequence modeling. The Bi-GRU consists of two opposite direction GRU units, which traverse the spatial sequence from left to right and right to left, respectively, and can capture the global semantic correlation between the previous and subsequent states. The output of the Bi-GRU is an intermediate representation encoding S emb = Bi-GRU(S ctx ), which contains the sequential semantics, structural features, and contextual information of the image in the spatial dimension, and helps to achieve more accurate feature discrimination and target recognition in the subsequent classification module.

[0067] The classification decision module is configured to receive the intermediate representation encoding S emb , calculate a probability distribution P cls using a nonlinear mapping function, output a multi-class label, and calculate an error term ε cls based on the probability distribution P gt and the true label P cls .

[0068] The classification decision module mainly consists of a mapping function construction unit and a multi-label decoding unit, which aims to convert the deep feature vector S emb into explicit classification results and output supervised error information for subsequent optimization. First, the mapping function construction unit converts the intermediate representation encoding S emb into a classification probability distribution P cls based on a nonlinear mapping mechanism. The mapping function adopts a softmax structure, which generates the prediction probability of the corresponding class by performing exponential mapping and normalization calculation on the linear relationship between each class weight ω j and S emb . The specific mapping function formula is:

[0069]

[0070] where 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 classes set in the image classification task, for example, it can be 10 traffic signs, 20 industrial defect types, or 1000 object recognition labels. In this way, 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 clsThe final classification result is determined. Specifically, this unit first sorts the probability values ​​of each predicted category and takes the category corresponding to the highest probability as the recognition label for the current image. Furthermore, to achieve end-to-end optimization training of the model, the multi-label decoding unit is also responsible for calculating the supervision error information, specifically described by the cross-entropy loss function. Its formula is as follows:

[0071]

[0072] in, This represents the true label of the k-th class, with a value range of [0,1]. Let K represent the predicted probability of the system's current output for the k-th class, where K represents the total number of labels, typically equivalent to the total number of classification categories C. This loss function reflects the deviation between the model's current output and the true labels, serving as a key error signal for subsequent backpropagation training.

[0073] In practical deployment, this classification decision module can support different types of image recognition tasks, such as identity classification in face recognition, defect localization and classification in industrial vision inspection, or lesion classification in medical images. By leveraging 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 possesses strong discriminative ability and scalability. Furthermore, the module employs weight normalization constraints... This also ensures that the model maintains a balanced distribution among the categories during the classification process, avoiding the bias problem caused by class imbalance.

[0074] An adaptive optimization module is used to optimize based on the error term ε. cls Update the weights of each learnable parameter in the system.

[0075] In the adaptive optimization module, the parameter update unit specifically relies on the supervision error term ε output by the classification decision module. cls Combined with adaptive learning rate η t With momentum factor μ, an iterative update rule with nonlinear characteristics is executed. First, for the currently learnable parameter ω in each training epoch... t The system will take the first derivative of the gradient of the parameter under the current error context to obtain the error term ε. cls For weight ω t The first-order partial derivative, i.e. Subsequently, a momentum factor μ∈[0,1] is introduced to fuse the weight update direction of the previous round with the current gradient direction. This aims to use historical information to smooth the update path, thereby avoiding parameter oscillations in the local minimum region and ensuring the continuity and stability of the optimization process.

[0076] In terms of learning rate, the embodiment does not adopt a fixed learning rate strategy, but introduces an adaptive learning rate t , which is dynamically adjusted according to the current gradient norm, and is defined as:

[0077]

[0078] wherein η0 is an initial learning rate, which is set as 0.2 by default in the system. The formula realizes effective suppression of gradient oscillation, that is, the step size is automatically reduced when the gradient is large, preventing excessive update; when the gradient is small, the step size is increased, accelerating the convergence speed, thereby improving the overall training efficiency.

[0079] The final parameter update formula is as follows:

[0080]

[0081] Here, through the weighted combination of the momentum term μ·ω t and the dynamic learning term , the model weight update path not only retains the consistency of the historical optimization direction, but also has the dynamic response ability to the current error. When the model error is large in the initial stage, η t is small, and the update step size is controlled, effectively avoiding overfitting; as the error gradually converges, the gradient term decreases, η t increases, accelerating the convergence speed.

[0082] It should be noted that, in the case of no conflict, the embodiments in the present application and the features and technical solutions in the embodiments can be combined with each other.

[0083] Obviously, the above-described embodiments are only some of the embodiments of the present application, not all the embodiments, and the preferred embodiments of the present application are given in the drawings, but do not limit the patent scope of the present application. The present application can be realized in many different forms, and conversely, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing specific embodiments, or make equivalent replacements to some technical features. Any equivalent structure made by using the contents of the present application specification and drawings, directly or indirectly applied to other related technical fields, is also within the patent protection scope of the present application.

Claims

1. A deep learning-based intelligent image recognition and classification system, characterized in that, Includes the following modules: Image preprocessing module, used to receive raw image data I raw The image data is then normalized, enhanced, and its resolution adjusted to output standardized image data I. std ; The feature extraction module is used to receive standardized image data I. std Extracting local texture features F based on multi-scale convolution kernels tex and edge features F edg To form a fused feature vector F ext The feature extraction module includes: Multi-scale convolutional units employ a set of scale-variable convolutional kernels. For standardized image data I std Extracting local texture features F tex The calculation formula is as follows: in, The standard ReLU activation function is used, and * represents the convolution operation; Edge enhancement unit, using Laplacian edge operator K lap Perform edge feature F edg Extraction, the Laplacian edge operator and with local texture features F tex The fused feature vector F is formed by splicing the features together. ext : F ext =Concat(F tex ,F edg ); Among them, F edg =K lap *I std ; The semantic modeling module is used to receive the fused feature vector F. ext The context-related semantic tensor S is extracted through a dual-channel attention network. ctx And generate intermediate representation encoding S emb ; The classification decision module is used to receive the intermediate representation code S. emb The probability distribution P is calculated using a nonlinear mapping function. cls To achieve multi-category label output, and based on probability distribution P cls and real label P gt Calculate the error term ε cls ; An adaptive optimization module is used to optimize based on the error term ε. cls The model gradient is calculated, and the model parameters are nonlinearly updated and optimized by combining the adaptive learning rate and momentum coefficient. The adaptive optimization module includes: The parameter update unit is used to update the error term ε output by the classification decision module. cls For the learnable weight parameters ω in the model t Perform iterative nonlinear optimization updates, combined with an adaptive learning rate η t The update rule for the momentum factor μ is as follows: Where, ω t ω represents the learned weights in the corresponding model. t+1 For the learned weights in the updated model The error term ε cls For weight ω t The first-order partial derivative, the momentum factor μ is in the range [0,1], η t The learning rate, dynamically adjusted based on the root mean square of the error gradient, is defined as: η0 is the initial learning rate, with a default value of 0.

2.

2. The intelligent image recognition and classification system based on deep learning according to claim 1, characterized in that, The image preprocessing module includes: Image enhancement unit, used to enhance the original image data I raw The adaptive correction enhancement algorithm is executed, and the enhancement expression formula is: Where, γ val The gamma correction exponent, γ val =1+0.25σ roi , σ roi The formula for calculating the luminance variance of the ROI region in an image is: L i Let L be the brightness value of the i-th pixel. roi is the average brightness within the ROI region of the image, and N is the number of pixels within the ROI region of the image; The resolution resampling unit uses a bilinear interpolation function to normalize the image to a uniform input size H. std ×W std This facilitates unified processing in subsequent operations.

3. The intelligent image recognition and classification system based on deep learning according to claim 1, characterized in that, The semantic modeling module includes: The channel attention modeling unit uses an SE structure to fuse the feature vector F. est Channel weighted generation of channel attention vector ω ch : Where σ is the Sigmoid activation function. The function is a standard ReLU activation function, where ω1 and ω2 are the learnable weights for the current channel, and satisfy ω1 + ω2 = 1. F ext To fuse feature vectors, H is the height of the image and W is the width of the image; The spatial context modeling unit employs a position-sensitive attention mechanism, which integrates the channel attention vector ω. ch Acting on the fused feature vector F ext Construct the context semantic tensor S ctx And a bidirectional gated loop function is used for semantic temporal modeling to generate intermediate representation encoding S. emb : WITH emb =Bi-GRU(S ctx ); Bi-GRU is a standard bidirectional gated loop function.

4. The intelligent image recognition and classification system based on deep learning according to claim 1, characterized in that, The classification decision module includes: The mapping function building block defines the nonlinear classification mapping as a probability distribution P. cls : Where, ω j Let be the weight parameters of the j-th class, and satisfy . C represents the total number of categories, S emb Encoding 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, The true label value for the k-th class, ranging from 0 to 1. Let K be the predicted probability of the model for the k-th class, where K represents the total number of labels.

Images