Image recognition method based on deep learning

Abstract

The invention relates to the field of image recognition, in particular to an image recognition method based on deep learning. The method comprises the following steps: collecting a historical image sample and a corresponding actual category and an actual category label; performing standardization processing and fast Fourier transform on the historical image samples to obtain a frequency domain feature matrix, and constructing a multi-stage feature fusion depth model; processing the frequency domain characteristic matrix to obtain a composite matrix; performing feature screening on the composite matrix to obtain a sparse feature coefficient vector, calculating a fuzzy membership degree, and outputting a prediction category; and based on the prediction category, optimizing the multi-stage feature fusion depth model, performing image recognition, and outputting a classification result. The method solves the problems that a current deep learning image recognition method is high in dependence on a large-scale data set and high in model construction cost; self-adaptability and flexibility are insufficient, and multi-scale information and complex modes in an image are difficult to fully capture; and the problems of over-fitting risk, model stability and recognition precision reduction exist.

Description

Technical Field

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

[0002] Current image recognition technology mainly relies on deep learning methods, such as convolutional neural networks (CNN) and deep neural networks (DNN). The method uses a large amount of data for supervised learning to extract multi-level features in the image, thereby achieving efficient image classification and recognition. In practice, image recognition technology has been widely used in various fields, such as medical image analysis, object detection in autonomous driving, security monitoring, etc. The above technology performs well in processing large-scale data sets and complex image tasks, and can achieve high-precision recognition results in a shorter time. However, the above technology usually relies on a large amount of labeled data, and the model complexity is high, and the computational and storage costs are high.

[0003] Although the current deep learning image recognition methods have made significant progress, there are still some problems to be solved: First, the current deep learning image recognition methods are highly dependent on large-scale data sets, and require a large amount of labeled data and computing resources during the training process, resulting in high model construction costs and difficulty in applying to resource-limited environments; second, they lack adaptability and flexibility in the feature extraction and fusion process, making it difficult to fully capture the multi-scale information and complex patterns in the image; in addition, there is a risk of overfitting, especially when facing data with high uncertainty or complexity, the stability and recognition accuracy of the model may be significantly reduced. In summary, the current deep learning image recognition methods still have a lot of room for improvement in processing complex image data, improving the robustness of the model, and reducing computational overhead, and urgently need to be improved through a new algorithm and processing technology. Summary of the invention

[0004] The present invention provides an image recognition method based on deep learning to solve the problems that the current deep learning image recognition method has a strong dependence on large-scale data sets, requires a large amount of labeled data and computing resources during the training process, resulting in high model construction costs and difficulty in application in resource-limited environments; lacks adaptability and flexibility in the feature extraction and fusion process, making it difficult to fully capture multi-scale information and complex patterns in images; and has the risk of overfitting, and when faced with data with high uncertainty or complexity, the stability of the model and the recognition accuracy are reduced.

[0005] The present invention provides an image recognition method based on deep learning, which specifically includes the following technical solutions:

[0006] A deep learning-based image recognition method comprises the following steps:

[0007] S1: Collect historical image samples and the actual categories and actual category labels corresponding to the historical image samples; perform standardization and fast Fourier transform on the historical image samples to obtain the frequency domain feature matrix; build a multi-stage feature fusion deep model based on the frequency domain feature matrix; introduce an adaptive geometric feature fusion algorithm to process the frequency domain feature matrix to obtain a synthetic matrix;

[0008] S2: Introduce a sparse feature screening mechanism to perform feature screening on the synthetic matrix to obtain a sparse feature coefficient vector; based on the sparse feature coefficient vector, calculate the fuzzy membership and select the actual category with the largest fuzzy membership as the predicted category; based on the predicted category, optimize the multi-stage feature fusion deep model; based on the optimized multi-stage feature fusion deep model, perform image recognition and output the classification results.

[0009] Preferably, the S1 specifically includes:

[0010] The multi-stage feature fusion deep model includes a feature fusion layer and a feature classification layer; in the feature fusion layer, an adaptive geometric feature fusion algorithm is introduced to perform multi-level feature fusion on the frequency domain feature matrix.

[0011] Preferably, the S1 specifically includes:

[0012] In the process of implementing the adaptive geometric feature fusion algorithm, geometric transformation and nonlinear adjustment are introduced based on the frequency domain feature matrix to calculate the geometric transformation feature matrix.

[0013] Preferably, the S1 specifically includes:

[0014] Through the high-dimensional embedding function, the linear embedding weight matrix and the nonlinear embedding weight matrix are introduced, the feature embedding mapping is performed on the geometric transformation feature matrix, and the geometric transformation feature matrix is ​​projected into the high-dimensional feature space to obtain the embedded feature matrix.

[0015] Preferably, the S1 specifically includes:

[0016] The embedded feature matrix and the geometric transformation feature matrix are mapped to different feature spaces through a multi-layer perceptron for feature decomposition, and the embedded feature matrices and geometric transformation feature matrices of different feature layers are obtained; the embedded feature matrices and geometric transformation feature matrices of different feature layers are feature fused through the hierarchical feature geometric synthesis formula to obtain a synthesis matrix; the hierarchical feature geometric synthesis formula is as follows:

[0017]

[0018] Among them, C i is a synthetic matrix, which represents the feature fusion result of the i-th historical image sample; is the total number of feature layers; l is the feature layer index; γl is the feature layer weight coefficient of the lth feature layer; M i，l is the embedding feature matrix of the lth feature layer in the i-th historical image sample; J is the number of interactive feature layers; j is the interactive feature layer index; δ lj is the interaction weight coefficient between the lth feature layer and the jth feature layer; G i，l is the geometric transformation feature matrix of the lth feature layer in the i-th historical image sample; G i，j is the geometric transformation feature matrix of the feature layer in the i-th historical image sample; ∈ is a positive constant used to prevent division by zero errors; ||·||2 is the L2 norm; ||·|| F is the Frobenius norm; T represents the transpose operation.

[0019] Preferably, the S2 specifically includes:

[0020] A sparse feature screening mechanism is introduced in the feature classification layer. Based on the synthetic matrix, the optimization objective function is minimized to obtain a sparse feature coefficient vector. The specific formula is as follows:

[0021]

[0022] Among them, A i is the sparse feature coefficient vector of the i-th historical image sample; D is the sparse dictionary matrix; λ1 is the L1 regularization parameter; λ2 is the Laplace regularization parameter; λ3 is the adjustment parameter of the singular value adjustment term; β is the weight adjustment parameter; L is the Laplace matrix; I is the identity matrix; ||·||1 is the L1 norm; Tr(·) is the matrix trace operation; σ max (·) is the maximum singular value operation.

[0023] Preferably, the S2 specifically includes:

[0024] Based on the sparse feature coefficient vector, fuzzy adjustment parameters and logarithmic adjustment items are introduced to calculate the fuzzy membership. Based on the fuzzy membership, the probability and uncertainty of historical image samples belonging to different categories are determined. The actual category with the largest fuzzy membership is selected as the predicted category of the historical image samples.

[0025] Preferably, the S2 specifically includes:

[0026] The predicted category is compared with the actual category label, the performance index of the multi-stage feature fusion deep model is calculated, and the performance thresholds are set respectively; when the performance index is lower than the corresponding performance threshold, the multi-stage feature fusion deep model is optimized using the back propagation algorithm and the gradient descent method; when the performance indicators all reach or exceed the performance threshold, or when the rate of change of the performance indicators is lower than the preset change threshold, the iterative optimization process is terminated; the optimized multi-stage feature fusion deep model is deployed in actual applications to receive real-time images, and the real-time images are processed using the above steps, image recognition is performed and classification results are output; the actual category labels identified are superimposed on the real-time images according to the classification results, and subsequent operations are performed.

[0027] The beneficial effects of the technical solution of the present invention are:

[0028] 1. By introducing an adaptive geometric feature fusion algorithm, the present invention can perform geometric transformation and nonlinear adjustment on the frequency domain feature matrix of historical images, significantly enhancing the discrimination and expression ability of frequency domain features. By combining the adaptive adjustment of the geometric transformation matrix and the nonlinear mapping matrix, the frequency domain features can be adaptively projected and expressed in different dimensions. At the same time, high-dimensional embedding functions and multi-layer perceptrons are used to realize multi-scale decomposition and fusion of features, thereby effectively capturing complex image patterns and improving the accuracy and robustness of image recognition.

[0029] 2. By using a sparse feature screening mechanism, the present invention can perform sparse feature screening on the fused synthetic matrix, effectively reduce the redundant features of the multi-stage feature fusion deep model, enhance the sparsity and robustness of the multi-stage feature fusion deep model, and reduce the risk of overfitting. The sparse feature screening mechanism combines reconstruction error, sparsity constraint, nonlinear interaction relationship between features and singular value adjustment, and uses matrix logarithm terms to capture complex nonlinear interactions. At the same time, the singular value adjustment term is used to enhance the sensitivity to eigenvalue differences, thereby improving the accuracy and efficiency of image recognition.

[0030] 3. The present invention introduces a fuzzy membership function, which calculates the fuzzy membership of the matching degree between sample features and actual categories to determine the probability and uncertainty of historical image samples belonging to different categories. The function uses fuzzy adjustment parameters and logarithmic adjustment items to enhance the differential characterization of image features, improve the sensitivity and stability of fuzzy membership calculation, and improve classification capabilities, so that the multi-stage feature fusion deep model shows stronger robustness and accuracy when processing uncertainty and complex data patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of an image recognition method based on deep learning described in the present invention. DETAILED DESCRIPTION

[0032] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0034] The following is a detailed description of a specific scheme of an image recognition method based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0035] Refer to the attached Figure 1 , which shows a flow chart of an image recognition method based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0036] S1: Collect historical image samples and the actual categories and actual category labels corresponding to the historical image samples; perform standardization and fast Fourier transform on the historical image samples to obtain the frequency domain feature matrix; build a multi-stage feature fusion deep model based on the frequency domain feature matrix; introduce an adaptive geometric feature fusion algorithm to process the frequency domain feature matrix to obtain a synthetic matrix;

[0037] Collect historical image samples and their corresponding actual category labels from databases or public datasets. The actual category is the real classification (such as disease type, object type) to which the historical image samples belong in the real world, and the actual category label is a symbol or number (such as a number or text) used to represent the actual category.

[0038] The historical image samples are standardized (noise removal, uniform size and grayscale) to obtain a standardized image matrix; the standardized image matrix is ​​converted to the frequency domain through fast Fourier transform (FFT) to obtain a frequency domain feature matrix, which contains the frequency domain features of the image (amplitude at different frequencies). Standardization and fast Fourier transform are existing technologies and will not be described in detail here.

[0039] Based on the frequency domain feature matrix, a multi-stage feature fusion deep model (Multi-Stage Feature Fusion Deep Model, MFFDM) is constructed; the multi-stage feature fusion deep model includes a feature fusion layer and a feature classification layer. The feature fusion layer applies an adaptive geometric feature fusion algorithm to the input frequency domain feature matrix to achieve multi-level feature fusion; the feature classification layer utilizes sparse feature screening to improve the image feature discrimination, calculates the fuzzy membership, and selects the actual category with the largest fuzzy membership as the predicted category of the historical image sample, and outputs the predicted category of the historical image sample as the classification result.

[0040] The adaptive geometric feature fusion algorithm combines geometric transformation and nonlinear adjustment to optimize the frequency domain feature matrix, and realizes multi-level feature fusion through high-dimensional embedding function and multi-layer perceptron decomposition to obtain a synthetic matrix; the specific process is as follows:

[0041] First, based on the frequency domain feature matrix, geometric transformation and nonlinear adjustment are introduced to calculate the geometric transformation feature matrix. The specific formula is as follows:

[0042]

[0043] Among them, G i is the geometric transformation feature matrix of the i-th historical image sample; is a geometric transformation matrix, which is used to project the frequency domain feature matrix to enhance the distinguishability of frequency domain features; F i is the frequency domain feature matrix of the i-th historical image sample; i is the sample index; is the mapping bias vector, which is used to translate and adjust the result after the projection of the frequency domain feature matrix to improve stability; α is the nonlinear adjustment coefficient, which is used to control the influence of the nonlinear term and is determined by the experimental tuning method; tanh is the hyperbolic tangent function, which is used to enhance the nonlinear expression ability of the frequency domain features; H is the nonlinear mapping matrix, which is used to introduce the nonlinear interaction relationship between the frequency domain features; T represents the transposition operation;

[0044] The initial values ​​of the geometric transformation matrix, mapping bias vector and nonlinear mapping matrix are obtained by random initialization method, and then adjusted and optimized by back propagation algorithm and gradient descent method in the training process of multi-stage feature fusion deep model. Back propagation algorithm and gradient descent method are prior art and will not be described in detail here;

[0045] Furthermore, the geometric transformation feature matrix is ​​mapped by feature embedding through a high-dimensional embedding function, and the geometric transformation feature matrix is ​​projected into a high-dimensional feature space to obtain an embedded feature matrix; the formula of the high-dimensional embedding function is as follows:

[0046]

[0047] Among them, M i is the embedding feature matrix of the i-th historical image sample; ReLU is the element-by-element activation function; U is the linear embedding weight matrix for linear transformation; b is the embedding bias vector for adjusting the center position of the feature embedding map; V is the nonlinear embedding weight matrix for nonlinear transformation;

[0048] The initial values ​​of the linear embedding weight matrix, the embedding bias vector, and the nonlinear embedding weight matrix are calculated by a random initialization method, and are subsequently adjusted and optimized during the training process of the multi-stage feature fusion deep model by a back propagation algorithm and a gradient descent method;

[0049] Next, a multi-layer perceptron is used to map the embedding feature matrix and the geometric transformation feature matrix to different feature spaces for feature decomposition, and the embedding feature matrix and the geometric transformation feature matrix of different feature layers are obtained, which are used to represent the multi-scale information and complex patterns of the embedding feature matrix and the geometric transformation feature matrix. For example, the low-frequency feature layer represents the overall structure and contour information of the image, and the high-frequency feature layer represents the details and texture information in the image. When designing the multi-layer perceptron, the total number of feature layers of the embedding feature matrix and the geometric transformation feature matrix remains the same. The multi-layer perceptron is a prior art and will not be described in detail here.

[0050] Furthermore, the embedded feature matrices and geometric transformation feature matrices of different feature layers are fused through the hierarchical feature geometric synthesis formula to obtain a synthesis matrix; the hierarchical feature geometric synthesis formula is as follows:

[0051]

[0052] Among them, C i is the synthesis matrix, which is the feature fusion result of the i-th historical image sample; is the total number of feature layers, which is set according to the specific implementation scenario; l is the feature layer index; γ l is the feature layer weight coefficient of the kth feature layer, indicating the weight of the lth feature layer in feature fusion; M i，l is the embedded feature matrix of the lth feature layer in the i-th historical image sample; J is the number of interactive feature layers, which indicates the number of feature layers selected for combination during feature fusion. It is obtained through experimental tuning, and the value of the number of interactive feature layers is less than or equal to the total number of feature layers; j is the interactive feature layer index; δ lj is the interaction weight coefficient between the lth feature layer and the jth feature layer, indicating the interaction strength between the feature layers; G i，l is the geometric transformation feature matrix of the lth feature layer in the i-th historical image sample; G i，jis the geometric transformation feature matrix of the feature layer in the i-th historical image sample; ∈ is a positive constant used to prevent division by zero errors; ||·||2 is the L2 norm; ||·|| F is the Frobenius norm;

[0053] The initial values ​​of the feature layer weight coefficient and the interaction weight coefficient are calculated by a random initialization method, and are subsequently adjusted and optimized step by step using a back propagation algorithm and a gradient descent method during the training process of the multi-stage feature fusion deep model.

[0054] S2: Introduce a sparse feature screening mechanism to perform feature screening on the synthetic matrix to obtain a sparse feature coefficient vector; based on the sparse feature coefficient vector, calculate the fuzzy membership and select the actual category with the largest fuzzy membership as the predicted category; based on the predicted category, optimize the multi-stage feature fusion deep model; based on the optimized multi-stage feature fusion deep model, perform image recognition and output the classification results.

[0055] The feature classification layer of the multi-stage feature fusion deep model uses the sparse feature screening (SFS) mechanism to screen the synthetic matrix to enhance the sparsity and robustness of the multi-stage feature fusion deep model, reduce the risk of overfitting, and improve the accuracy and efficiency of image recognition. The optimization goal of the sparse feature screening mechanism combines reconstruction error, sparsity constraints, nonlinear interactions between features, and singular value adjustment. The logarithmic term is introduced to capture complex nonlinear interactions, and the singular value adjustment term is used to enhance the sensitivity to eigenvalue differences. The specific implementation process of the sparse feature screening mechanism is as follows:

[0056] First, by minimizing the optimization objective function, the sparse feature coefficient vector is obtained. The specific formula is as follows:

[0057]

[0058] Among them, A i is the sparse feature coefficient vector of the i-th historical image sample; D is the sparse dictionary matrix, which is the basis for feature representation; represents the reconstruction error; λ1||A i ||1 represents the sparsity constraint; λ1 is the L1 regularization parameter, which is used to enhance the sparsity of the multi-stage feature fusion deep model; Represents the nonlinear interaction between features; λ2 is the Laplace regularization parameter, which is used for smooth feature selection; λ3 is the adjustment parameter of the singular value adjustment term, which is used to control the maximum singular value adjustment term The weight in the optimization objective function; β is the weight adjustment parameter, which is used to balance the influence of the last item (the difference item based on singular values) in the optimization objective function on the optimization objective function; D, λ1, λ2, λ3, and β are set according to the specific implementation scenario using expert experience; L is the Laplace matrix, which is used to maintain the local smoothness between the sparse feature coefficient vectors; I is the identity matrix; ||·||1 is the L1 norm, which is used to calculate the sparsity of the sparse feature coefficient vector; Tr(·) is the trace operation of the matrix; σ max (·) is the maximum singular value operation;

[0059] Next, based on the sparse feature coefficient vector, the fuzzy membership is calculated; based on the fuzzy membership, the probability and uncertainty of the historical image samples belonging to different categories are determined. The specific formula of the fuzzy membership function is as follows:

[0060]

[0061] Among them, μ ic is the fuzzy membership of the i-th historical image sample to the actual category c, indicating the matching degree of the i-th historical image sample to the actual category c; η is the fuzziness adjustment parameter, which is used to control the steepness of the fuzzy membership function; w c is the category weight vector of the actual category c, which maps the input sparse feature coefficient vector to the feature space of the actual category c and is used to calculate the fuzzy membership of the i-th historical image sample to the actual category c; c is the actual category index; is the category decision bias of the actual category c, expressed in scalar form, used to adjust the decision boundary of the actual category; θ is the weight factor of the fuzziness adjustment, which is used to adjust the influence of the logarithmic adjustment term and is obtained through experimental tuning; is a logarithmic adjustment term used to enhance The difference of , smoothing the value range, thus improving the sensitivity and stability of fuzzy membership calculation;

[0062] The initial values ​​of the category weight vector and category decision bias are calculated by a random initialization method, and are subsequently adjusted and optimized step by step using a back propagation algorithm and a gradient descent method during the training process of the multi-stage feature fusion deep model;

[0063] Select the actual category with the largest fuzzy membership as the predicted category of the historical image sample, that is, the classification result of the multi-stage feature fusion deep model. Compare the predicted category with the actual category label, calculate the performance index of the multi-stage feature fusion deep model, the performance index includes accuracy, recall rate, F1 score, and set performance thresholds respectively, the performance thresholds include accuracy threshold, recall rate threshold, F1 score threshold, if there is a performance index lower than the corresponding performance threshold, use the back propagation algorithm and gradient descent method to optimize the multi-stage feature fusion deep model;

[0064] When the performance indicators (accuracy, recall rate, F1 score) all reach or exceed the preset performance threshold, or when the change rate of the performance indicators after a certain number of iterations is lower than the set change threshold, the iterative optimization process is terminated; the optimized multi-stage feature fusion deep model is deployed to actual applications (such as servers, cloud platforms, embedded devices or edge computing devices) to receive real-time images, and the real-time images are processed using the above steps to perform image recognition and output classification results. The actual category labels identified are superimposed on the real-time images according to the classification results, and subsequent operations (such as alarms, data storage, and further analysis) are performed. The specific subsequent operations are set according to the specific implementation scenarios.

[0065] Continue to monitor the performance of the multi-stage feature fusion deep model in actual applications, including performance indicators such as accuracy, recall, and F1 score. If it is found that the performance indicators of the multi-stage feature fusion deep model are lower than the performance thresholds set above, use the back propagation algorithm and gradient descent method to optimize the multi-stage feature fusion deep model.

[0066] In summary, a deep learning-based image recognition method is completed.

[0067] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0068] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. An image recognition method based on deep learning, characterized in that: The following steps are involved: S1: Collect historical image samples and the actual categories and actual category labels corresponding to the historical image samples; perform standardization and fast Fourier transform on the historical image samples to obtain the frequency domain feature matrix; Based on the frequency domain feature matrix, a multi-stage feature fusion deep model is constructed; Introduction Adaptive geometric feature fusion algorithm processes the frequency domain feature matrix to obtain a synthetic matrix; S2: Introduce a sparse feature screening mechanism to perform feature screening on the synthetic matrix and obtain a sparse feature coefficient vector; Based on the sparse feature coefficient vector, the fuzzy membership is calculated, and the actual category with the largest fuzzy membership is selected as the predicted category; based on the predicted category, the multi-stage feature fusion deep model is optimized; based on the optimized multi-stage feature fusion deep model, image recognition is performed and the classification results are output.

2. The image recognition method based on deep learning according to claim 1, characterized in that: The S1 specifically includes: The multi-stage feature fusion deep model includes a feature fusion layer and a feature classification layer; in the feature fusion layer, an adaptive geometric feature fusion algorithm is introduced to perform multi-level feature fusion on the frequency domain feature matrix.

3. The image recognition method based on deep learning according to claim 2, characterized in that: The S1 specifically includes: In the process of implementing the adaptive geometric feature fusion algorithm, geometric transformation and nonlinear adjustment are introduced based on the frequency domain feature matrix to calculate the geometric transformation feature matrix.

4. The image recognition method based on deep learning according to claim 3, characterized in that: The S1 specifically includes: Through the high-dimensional embedding function, the linear embedding weight matrix and the nonlinear embedding weight matrix are introduced, the feature embedding mapping is performed on the geometric transformation feature matrix, and the geometric transformation feature matrix is ​​projected into the high-dimensional feature space to obtain the embedded feature matrix.

5. The image recognition method based on deep learning according to claim 4, characterized in that: The S1 specifically includes: The embedded feature matrix and the geometric transformation feature matrix are mapped to different feature spaces through a multi-layer perceptron for feature decomposition, and the embedded feature matrices and geometric transformation feature matrices of different feature layers are obtained; the embedded feature matrices and geometric transformation feature matrices of different feature layers are feature fused through the hierarchical feature geometric synthesis formula to obtain a synthesis matrix; the hierarchical feature geometric synthesis formula is as follows: Among them, C i is a synthetic matrix, which represents the feature fusion result of the i-th historical image sample; is the total number of feature layers; l is the feature layer index; γ l is the feature layer weight coefficient of the lth feature layer; M i，l is the embedding feature matrix of the lth feature layer in the i-th historical image sample; J is the number of interactive feature layers; j is the interactive feature layer index; δ lj is the interaction weight coefficient between the lth feature layer and the jth feature layer; G i，l is the geometric transformation feature matrix of the lth feature layer in the i-th historical image sample; G i，j is the geometric transformation feature matrix of the feature layer in the i-th historical image sample; ∈ is a positive constant used to prevent division by zero errors; ||·||2 is the L2 norm; ||·|| F is the Frobenius norm; T represents the transpose operation.

6. The image recognition method based on deep learning according to claim 5, characterized in that: The S2 specifically includes: A sparse feature screening mechanism is introduced in the feature classification layer. Based on the synthetic matrix, the optimization objective function is minimized to obtain a sparse feature coefficient vector. The specific formula is as follows: Among them, A i is the sparse feature coefficient vector of the i-th historical image sample; D is the sparse dictionary matrix; λ1 is the L1 regularization parameter; λ2 is the Laplace regularization parameter; λ3 is the adjustment parameter of the singular value adjustment term; β is the weight adjustment parameter; L is the Laplace matrix; I is the identity matrix; ||·||1 is the L1 norm; Tr(·) is the matrix trace operation; σ max (·) is the maximum singular value operation.

7. The image recognition method based on deep learning according to claim 6, characterized in that: The S2 specifically includes: Based on the sparse feature coefficient vector, fuzzy adjustment parameters and logarithmic adjustment items are introduced to calculate the fuzzy membership. Based on the fuzzy membership, the probability and uncertainty of historical image samples belonging to different categories are determined. The actual category with the largest fuzzy membership is selected as the predicted category of the historical image samples.

8. The image recognition method based on deep learning according to claim 7, characterized in that: The S2 specifically includes: The predicted category is compared with the actual category label, the performance index of the multi-stage feature fusion deep model is calculated, and the performance thresholds are set respectively; when the performance index is lower than the corresponding performance threshold, the multi-stage feature fusion deep model is optimized using the back propagation algorithm and the gradient descent method; when the performance indicators all reach or exceed the performance threshold, or when the rate of change of the performance indicators is lower than the preset change threshold, the iterative optimization process is terminated; the optimized multi-stage feature fusion deep model is deployed in actual applications to receive real-time images, and the real-time images are processed using the above steps, image recognition is performed and classification results are output; the actual category labels identified are superimposed on the real-time images according to the classification results, and subsequent operations are performed.

Images