Image classification method and device based on combined hierarchical development network, and medium
By adopting a joint hierarchical development network in image classification, combining feature extraction and development attention modules, the problem of low perception of development network tasks is solved, and higher image classification accuracy and semantic information capture ability are achieved, which improves the network's Robustness and scalability.
Patent Information
- Application Number
- CN202410959500.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-12-03
AI Technical Summary
Existing developmental networks have low task perception in image classification, difficult to capture high-level semantic information, and poor scalability and binding.
The image classification method based on a joint hierarchical development network is adopted, including feature extraction module, development attention module and task head module. Through feature extraction, development attention mechanism and task-driven top-down attention mechanism, combined with supervision information and gradient inversion Propagation algorithms, update parameters to improve task perception.
It improves the accuracy of image classification and the network's ability to capture high-level semantic information, and enhances the robustness and scalability of the network.
Smart Images

Figure CN119068229A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to an image classification method, device and medium based on a joint hierarchical developmental network. Background Art
[0002] The attention mechanism originated from cognitive research on the human brain. Since the scale that the human visual nervous system can process simultaneously is much smaller than its actual acceptance, a phenomenon of "selective attention" was discovered, that is, when observing a complex external environment, humans can autonomously and efficiently focus on the salient target area of interest and form a focus of attention, thereby maximizing the efficiency of visual information analysis under the premise of limited human brain processing resources. For computer vision tasks, this mechanism can be understood as an adaptive dynamic feature selection process, that is, for a specific input image feature map, differentiated attention weight values are given in specific dimensions according to the established task requirements, thereby achieving a division of importance between different features. With the support of this mechanism, the model simulates the perception and understanding of the human visual system to the greatest extent, so that it is more inclined to pay attention to some information in the image that assists judgment and ignores other irrelevant information.
[0003] With the deepening of cognitive research, people have found that the attention mechanism has a rich regulatory pathway, namely, the "bottom-up" attention driven by stimuli and the "top-down" attention driven by tasks. The former is driven by physical features and has a good perception of features; the latter relies on prior knowledge and has better task perception based on physical features. However, looking at the development of the attention mechanism, most of the existing attention mechanism methods mainly design the attention mechanism based on "methods with the ability to effectively extract key information", which is actually a feature-based "bottom-up" attention mechanism, lacking understanding, perception and flexibility of specific tasks.
[0004] Despite this, some researchers have explored task-based "top-down" attention mechanisms. The developmental network is a classic representative of this. However, due to the inherent shallow characteristics of the network and the difference from the parameter update mechanism of the gradient back-propagation algorithm, the network has poor scalability and associativity, and it is difficult to capture high-level semantic information, and the task perception ability cannot be well applied. Summary of the invention
[0005] The purpose of the present invention is to solve the problem of low perception ability of developmental network tasks. The present invention provides an image classification method, device and medium based on a joint hierarchical developmental network. A method for image classification based on a joint hierarchical developmental network, the specific steps include: S1. Obtain label data corresponding to the image dataset to be classified and divide the dataset; S2, build an image classification network based on the joint hierarchical developmental network and train it using the dataset; S3. Use the image classification network to process the images to be classified in the verification set to obtain image classification results.
[0006] Furthermore, the image classification network based on the joint hierarchical developmental network includes three parts: feature extraction module, developmental attention module and task head module; The feature extraction module is used to extract the initial features of different levels of the image. It uses a lightweight network Resnet-18, which has five stages and can extract initial features of different levels. The task header module is used to generate task results; The developmental attention module is based on the developmental network core, including the receptive layer, hidden layer and effect layer, which is used to select the initial features.
[0007] Furthermore, the specific steps of image classification network training based on the joint hierarchical developmental network are: S21, sending the training set data to the feature extraction module to obtain initial features; S22, using the initial features to train the developmental attention module and update parameters; S23, inputting the initial features into the task head module and the trained developmental attention module respectively, and combining the outputs of the two to obtain the image classification result; S24. Calculate the loss function and perform gradient back propagation update through the classification head module and the development attention module respectively, specifically: S241. Calculate the loss function, which includes two parts: supervision loss and consistency loss. The formula is as follows:
[0008] in, is the loss function; is the supervised loss, and the latter is the reconstruction loss; S242, retaining the activated neurons of the hidden layer and the weights connected thereto, and performing gradient calculation according to the processing method of the linear layer; S25, repeat S23-S24 until the loss drops to a threshold or the training round reaches an upper limit, then stop training; S26. Reset the parameters of the developmental attention module, repeat S22-S25 until the training is terminated, and use the final output of S23 as the image result.
[0009] Furthermore, the parameters that need to be updated in the developmental attention module include: The weight matrices of the perception layer, hidden layer, and effect layer are ( ), , , are all random matrices at the beginning of training; the age of neurons in the hidden layer , hidden layer neuron activation threshold and the age of effector neurons , are all 0 at the beginning of training; Furthermore, the specific steps of training the developmental attention module are: S221, the receptive area receives the initial features from the feature extraction module , the effect area receives the label data corresponding to the initial features in the training set , the deepest features The label data z is weighted and summed with the corresponding weight matrix to obtain the pre-activation value , the mathematical form is as follows:
[0010] S222, remove pre-activation values in the hidden layer The neurons with the maximum pre-activation value are smaller than the activation threshold of the hidden layer neurons, and the neurons corresponding to the maximum pre-activation value are used as the activation neurons of the hidden layer. At the same time, the sample label z is encoded into a one-hot code form z, and the neurons in the activation effect layer with the same index as the effective coding bit are activated; S223, completing the updating of the parameters in the weight matrix according to the leaf component analysis method, and updating the neuron age at the same time; The updating process of the parameters in the weight matrix is:
[0011]
[0012]
[0013] In the formula, represents the weight matrix at time t; are the retention rate and learning rate respectively; Indicates neuronal activation markers; is the neuronal age; Furthermore, the updating mechanism of neuron age is as follows: The age of the activated neurons in the hidden layer increases by one unit each time they are selected; the age of the neurons in the effect area is updated in an adaptive manner, specifically: When the sum of the ages of all neurons in the effect area is less than the upper limit, the age of the corresponding activated Z neuron will also increase by one unit; when it is greater than the upper limit, the connection strength between the corresponding activated Z neuron and the Y area is quantified as the neuron age. The mathematical expression is as follows:
[0014] In the formula, e is the upper age limit; is the count of all activated neurons; represents the activation flag of the jth neuron; S224. Repeat the training until the number of activated neurons in region Y in the developmental network reaches a set value. The training stops.
[0015] Furthermore, the activation threshold of the hidden layer neurons is adaptively updated as the training progresses. The specific update mechanism is as follows:
[0016] in, represents the neuron activation threshold at time t; It represents the activation value of the neuron that finally wins after competitive screening. represents the activation value of the neuron with the largest activation value in pre-activation, is a threshold control coefficient.
[0017] Furthermore, in step S22, when facing multiple samples, each parameter is updated in the following manner: The weight matrix is updated as follows: =
[0018] =
[0019] Where N is the number of samples; Represents the activation value of the hidden layer neuron during the i-th sample training; represents the weight matrix of the i-th sample updated according to the leaf component analysis method; The hidden area activation threshold generated by N samples is updated as follows:
[0020] in, represents the hidden layer neuron activation threshold during the i-th sample training; For N samples, the neuron age is updated as follows:
[0021] N In the formula, Represents the age of neurons in the effect layer during the training of the i-th sample.
[0022] Further, step S23 is specifically as follows: S231, the deepest feature x in the initial features kThe corresponding feature vector R1 is generated in the input task head; S232, input the initial features of different levels into the developmental attention module X area, and compare them with the weight matrix trained in S2 Multiply them, and take the weighted sum of all the multiplication results as the Y area activation value, and Multiply and output the corresponding eigenvector R2; S233. Normalize R1 and R2 respectively, fuse them to obtain a feature vector R, and take the category corresponding to the index of the maximum eigenvalue in R as the result of image classification.
[0023] A storage medium stores instructions and data for implementing an image classification method based on a joint hierarchical developmental network.
[0024] A computer device, characterized in that it includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an image classification method based on a joint hierarchical developmental network.
[0025] The beneficial effects brought about by the technical solution provided by the present invention are: the image classification network constructed based on the joint hierarchical developmental network of the present invention provides rich task perception capabilities for image classification and improves the accuracy of image classification; by introducing adaptive parameters and multi-sample processing mechanisms in the image classification network, the network's ability to capture high-level semantic information is improved, while the network's robustness and scalability are improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a flow chart of an image classification method based on a joint hierarchical developmental network in an embodiment of the present invention; Figure 2 is a flow chart of an image classification method based on a joint hierarchical developmental network in an embodiment of the present invention; Figure 3 is a schematic diagram of updating the development attention module in an embodiment of the present invention; Figure 4 is a schematic diagram of a training phase of an image classification method based on a joint hierarchical developmental network in an embodiment of the present invention; Figure 5 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION
[0027] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0028] The embodiments of the present invention provide an image classification method, device and storage medium based on a joint hierarchical developmental network.
[0029] Please refer to Figure 1 , Figure 1 It is a flowchart of an image classification method based on a joint hierarchical developmental network in an embodiment of the present invention.
[0030] An image classification method based on a joint hierarchical developmental network, the specific steps include: The first step is to obtain the image data set and label data to be classified and divide the data set.
[0031] In order to verify the effectiveness of the present invention, experiments were conducted on the mainstream image classification dataset Cifar-100.
[0032] CIFAR-100 is a medium-sized dataset consisting of 100 categories of objects commonly found in nature and daily life, including 60,000 color images with a pixel value of 32x32. Each category has 600 image samples, 500 for training and 100 for testing. Since the images in this dataset are all collected in the real world, there are common difficulties in classification tasks such as high noise, complex features, and large changes in object scale, so it is often used to verify the performance of classification algorithms.
[0033] In this experiment, color images are directly input into the network as image samples, and all categories are numbered without duplication, and the corresponding category numbers of the images are input into the network as label samples.
[0034] The second step is to build an image classification network based on the joint hierarchical developmental network and train it using the training set.
[0035] The image classification network based on the joint hierarchical developmental network consists of three parts: feature extraction module, developmental attention module and task head module: The feature extraction module is used to extract the initial features of different levels of the image. It uses a lightweight network Resnet-18, which has five stages and can extract initial features of different levels. The network has been pre-trained on ImageNet, which can reduce the convergence time to a certain extent.
[0036] The task header module is used to generate task results.
[0037] The developmental attention module is based on the developmental network core, including the receptive layer, hidden layer and effect layer, which is used to select the initial features.
[0038] Developmental Network (DN) is a lightweight brain-like network that naturally has a top-down attention mechanism. It is analogous to the various links of human information transmission function, and cleverly divides the network into three levels: perception layer, hidden layer, and effect layer. During the training phase, while the perception layer accepts natural image input, the effect layer also accepts the input of supervisory information to participate in calculation and learning. In the testing phase, the effect layer is transformed into the output end to produce the recognition result under the corresponding task. In this way, the supervisory information is naturally integrated into the algorithm, thereby achieving a good perception of the task.
[0039] The training steps of the image classification network based on the joint hierarchical developmental network are as follows: Figure 2 As shown, specifically: In the first step, the images to be classified in the training set are sent to the feature extraction module to obtain the initial features.
[0040] In the second step, the initial features are used to train the developmental attention module and update the parameters, such as Figure 3 shown.
[0041] The parameters that need to be updated include the weight matrices of the perception layer, hidden layer, and effect layer, which are ( ), , , are all random matrices at the beginning of training; is the age of neurons in the hidden layer, is the hidden layer neuron activation threshold, is the age of neurons in the effect layer, which is 0 at the beginning of training. At the same time, the number of neurons is preset to be 2-3 times the number of batches.
[0042] The specific steps of parameter updating are: Step 1: The receptive area receives the initial features from the feature extraction module , the effect area receives the label data of the initial features , the deepest features The label data z is weighted and summed with the corresponding weight matrix to obtain the pre-activation value , the mathematical form is as follows:
[0043] Step 2: Remove pre-activation values in the hidden layer The neurons whose value is less than the activation threshold of the hidden layer neurons are activated, and the neurons corresponding to the maximum pre-activation value are used as the activation neurons of the hidden layer. At the same time, the corresponding neurons of the effect layer are also activated.
[0044] Step 3: Update the parameters in the weight matrix according to the leaf component analysis method, and update the neuron age at the same time.
[0045] The updating process of the parameters in the weight matrix is:
[0046]
[0047]
[0048] In the formula, represents the weight matrix at time t; are the retention rate and learning rate, respectively.
[0049] The specific mechanism of neuronal age update is: The age of the activated neurons in the hidden layer increases by one unit each time they are selected; The age of neurons in the effector area is updated in an adaptive manner, specifically: When the sum of the ages of all neurons in the effect area is less than the upper limit, the age of the corresponding activated Z neuron will also increase by one unit; when it is greater than the upper limit, the connection strength between the corresponding activated Z neuron and the Y area is quantified as the neuron age. The mathematical expression is as follows:
[0050] In the formula, e is the upper age limit; is the count of all activated neurons; represents the i-th activated neuron.
[0051] The activation threshold of hidden layer neurons is adaptively updated as training progresses. The specific update mechanism is as follows:
[0052] in, represents the neuron activation threshold at time t; It represents the activation value of the neuron that finally wins after competitive screening. represents the activation value of the neuron with the largest activation value in pre-activation, is a threshold control coefficient.
[0053] Under this threshold mechanism, when the neuron corresponding to the maximum activation value in the pre-activation value loses in the competition, it will re-participate in learning, so that the threshold can be stable and appropriate during training; for other neurons, the threshold control coefficient is used to make them easier to activate, so as to prevent new neurons from being over-activated due to excessive thresholds in the early stages of training.
[0054] Step 4: Repeat the training until the number of activated neurons in region Y in the developmental network reaches the set value.
[0055] Since the initial features usually package samples into batches for training, when faced with multiple samples, the present invention proposes a multi-sample learning mechanism, and the parameter update method is as follows: The weight matrix is updated as follows: =
[0056] =
[0057] Where N is the number of samples; Represents the activation value of the hidden layer neuron during the i-th sample training; The hidden area activation threshold generated by N samples is updated as follows:
[0058] in, represents the hidden layer neuron activation threshold during the i-th sample training; For N samples, the neuron age is updated as follows:
[0059] N In the third step, the initial features are input into the task head module and the trained developmental attention module respectively, and the output of the two is combined to obtain the output of the network, which is: Step 1: Input the deepest feature xk in the initial features into the task head to generate the corresponding feature vector R1.
[0060] Step 2: Input the initial features at different levels into the developmental attention module X region and compare them with the trained weight matrix in S2 Multiply them, and take the weighted sum of all the multiplication results as the Y area activation value, and The multiplication outputs the corresponding eigenvector R2.
[0061] Step 3: Normalize R1 and R2 respectively, fuse them to get the feature vector R, and take the category corresponding to the index of the maximum eigenvalue in R as the result of image classification.
[0062] In step 4, the loss function is calculated and updated by gradient backpropagation through the classification head module and the developmental attention module respectively.
[0063] Step 1: Calculate the loss function, which includes supervision loss and consistency loss. The formula is as follows:
[0064] in, is the loss function; is a supervised loss, which is used for supervised learning of specific tasks, such as cross entropy loss in classification tasks; the latter is a reconstruction loss, hoping that the features of each level It can approach the inductive features learned by the developmental attention module, thereby making the generated initial features more in line with the needs of the developmental attention network.
[0065] Step 2: Since the developmental attention module uses leaf component analysis to update parameters, it is impossible to perform gradient calculation and back-propagation update. Therefore, a gradient calculation strategy based on developmental networks is proposed. By treating the connections between the X, Y, and Z regions as multi-layer perceptrons, only the activated neurons in the Y region and the weights connected to them are retained each time the gradient is calculated, and the gradient calculation is performed according to the processing method of the linear layer.
[0066] Step 5: Repeat steps 3 and 4 until the loss drops to the threshold or the training rounds reach the upper limit.
[0067] In step 6, reset the parameters of the developmental attention module and repeat steps 2 to 5 until the training is terminated. The final output of step 3 is used as the image classification result.
[0068] like Figure 4 As shown in the figure, the alternating training strategy adopted in the training phase, the second step is called the "feature selection phase", through the training of the developmental attention module, the corresponding features are aggregated for the task target to generate features with task attention; the third to fifth steps are called the "joint adjustment phase", through the optimization strategy of minimizing the loss function, the feedback and fusion of task attention are realized. The two stages are carried out alternately, on the one hand, it can continuously regulate the realization of task attention, and on the other hand, it can also alleviate the dependence of the developmental attention module on the stability of input data.
[0069] In this embodiment, since the feature extraction module features (backbone network) adopts the commonly used convolutional neural network Resnet-18, the developmental attention module uses the developmental neural network as the core algorithm. Therefore, the present invention is compared with the developmental neural network (DN), the attention developmental neural network (ADN), Resnet-18 (Resnet), Resnet-18 with feature redundancy loss (FRL), and the Squeeze-and-Excitaion Networks (SENet) based on the channel attention mechanism on the same data set. The classification results in the cifar-100 test set are shown in the following table: Table 1 Classification indicators of the cifar-100 dataset method Backbone network Accuracy DN - 0.305 ADN - 0.342 Resnet Resnet-18 0.761 FRL Resnet-18 0.766 SENet Resnet-18 0.774 The present invention Resnet-18 0.786 From the experimental results in Table 1, it can be seen that the proposed method is significantly better than the existing method, which proves the superiority of the proposed method.
[0070] See also Figure 5 , Figure 5 40 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a computer device 401, a processor 402 and a storage medium 403.
[0071] A computer device 401: The computer device 401 implements the image classification method based on a joint hierarchical developmental network.
[0072] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the image classification method based on the joint hierarchical developmental network.
[0073] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the image classification method based on the joint hierarchical developmental network.
[0074] The beneficial effects of the present invention are as follows: the image classification network constructed based on the joint hierarchical developmental network provides rich task perception capabilities for image classification and improves the accuracy of image classification; by introducing adaptive parameters and multi-sample processing mechanisms in the image classification network, the network's ability to capture high-level semantic information is improved, while the network's robustness and scalability are improved.
[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these 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.
Claims
1. An image classification method based on a joint hierarchical developmental network, characterized in that: The specific steps include: S1. Obtain the image data set to be classified and divide the data set; S2, obtain the label data corresponding to the training set data, and add it to the training set data; S3, build an image classification network based on the joint hierarchical developmental network and train it using the dataset; S4. Use the image classification network to process the image to be classified to obtain the image classification result.
2. The image classification method based on a joint hierarchical developmental network as claimed in claim 1, characterized in that: The image classification network based on the joint hierarchical developmental network comprises three parts: a feature extraction module, a development attention module and a task head module; The feature extraction module is used to extract initial features of different levels of the image; a lightweight network Resnet-18 is used, which has five stages and can extract initial features of different levels respectively; The task header module is used to generate task results; The developmental attention module is based on a developmental network core, including a receptive layer, a hidden layer and an effect layer, which is used to select initial features.
3. The image classification method based on a joint hierarchical developmental network as claimed in claim 2, characterized in that: The specific steps of the image classification network training based on the joint hierarchical developmental network are: S21, sending the training set data to the feature extraction module to obtain initial features; S22, using the initial features to train the developmental attention module and update parameters; S23, inputting the initial features into the task head module and the trained developmental attention module respectively, and combining the outputs of the two to obtain the image classification result; S24. Calculate the loss function and perform gradient back propagation update through the classification head module and the development attention module respectively, specifically: S241. Calculate the loss function, which includes two parts: supervision loss and consistency loss. The formula is as follows: in, is the loss function; is the supervised loss, and the latter is the reconstruction loss; S242, retaining the activated neurons of the hidden layer and the weights connected thereto, and performing gradient calculation according to the processing method of the linear layer; S25, repeat S23-S24 until the loss drops to a threshold or the training round reaches an upper limit, then stop training; S26. Reset the parameters of the developmental attention module, repeat S22-S25 until the training is terminated, and use the final output of S23 as the image result.
4. The image classification method based on a joint hierarchical developmental network as claimed in claim 3, characterized in that: The parameters that need to be updated in the developmental attention module include: The weight matrices of the perception layer, hidden layer, and effect layer are ( ), , , are all random matrices at the beginning of training; the age of neurons in the hidden layer , hidden layer neuron activation threshold and the age of effector neurons , are all 0 at the beginning of training; The specific steps of the training development attention module are: S221, the receptive area receives the initial features from the feature extraction module , the effect area receives the label data corresponding to the initial features in the training set , the deepest features The label data z is weighted and summed with the corresponding weight matrix to obtain the pre-activation value , the mathematical form is as follows: S222, remove pre-activation values in the hidden layer The neurons with the maximum pre-activation value are smaller than the activation threshold of the hidden layer neurons, and the neurons corresponding to the maximum pre-activation value are used as the activation neurons of the hidden layer. At the same time, the label data z is encoded into a one-hot code form to activate the neurons in the effect layer with the same encoding bit index; S223, completing the updating of the parameters in the weight matrix according to the leaf component analysis method, and updating the neuron age at the same time; The updating process of the parameters in the weight matrix is: In the formula, represents the weight matrix at time t; are the retention rate and learning rate respectively; Indicates neuronal activation markers; is the neuronal age; The specific mechanism of neuronal age update is: The age of the activated neurons in the hidden layer increases by one unit each time they are selected; the age of the neurons in the effect area is updated in an adaptive manner, specifically: When the sum of the ages of all neurons in the effect area is less than the upper limit, the age of the corresponding activated Z neuron will also increase by one unit; when it is greater than the upper limit, the connection strength between the corresponding activated Z neuron and the Y area is quantified as the neuron age. The mathematical expression is as follows: In the formula, e is the upper age limit; is the count of all activated neurons; represents the activation flag of the jth neuron; S224. Repeat the training until the number of activated neurons in region Y in the developmental network reaches a set value. The training stops.
5. The image classification method based on a joint hierarchical developmental network as claimed in claim 4, characterized in that: The hidden layer neuron activation threshold is adaptively updated as the training progresses, and the update mechanism is specifically as follows: in, represents the neuron activation threshold at time t; It represents the activation value of the neuron that finally wins after competitive screening. represents the activation value of the neuron with the largest activation value in pre-activation, is a threshold control coefficient.
6. The image classification method based on a joint hierarchical developmental network as claimed in claim 5, characterized in that: When facing multiple samples, step S22 updates the parameters in the following way: The weight matrix is updated as follows: = = Where N is the number of samples; Represents the activation value of the hidden layer neuron during the i-th sample training; represents the weight matrix of the i-th sample updated according to the leaf component analysis method; The hidden area activation threshold generated by N samples is updated as follows: in, represents the hidden layer neuron activation threshold during the i-th sample training; For N samples, the neuron age is updated as follows: N In the formula, Represents the age of neurons in the effect layer during the training of the i-th sample.
7. The image classification method based on a joint hierarchical developmental network as claimed in claim 6, characterized in that: Step S23 is specifically as follows: S231, the deepest feature x in the initial features k The corresponding feature vector R1 is generated in the input task head; S232, input the initial features of different levels into the developmental attention module X area, and compare them with the weight matrix trained in S2 Multiply them, and take the weighted sum of all the multiplication results as the Y area activation value, and Multiply and output the corresponding eigenvector R2; S233. Normalize R1 and R2 respectively, fuse them to obtain the feature vector R, and take the category corresponding to the index of the maximum eigenvalue in R as the result of image classification.
8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing an image classification method based on a joint hierarchical developmental network as described in any one of claims 1 to 7.
9. A computer device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement an image classification method based on a joint hierarchical developmental network as described in any one of claims 1 to 7.
