Public facility detection method based on convolutional neural network GoogleNet
By adopting an improved GoogleNet network in public facility identification, using two consecutive 3x3 convolutional layers and Leaky ReLU activation functions, combined with data standardization operations, the problems of low recognition accuracy and insufficient generalization capabilities are solved, and more efficient and stable public facility identification is achieved.
Patent Information
- Application Number
- CN202411556427.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-02-25
AI Technical Summary
The prior art has problems in the identification of public facilities with low recognition accuracy, insufficient generalization ability and overfitting risks, which are difficult to meet the needs of intelligent management.
The detection method based on the convolution neural network GoogleNet is adopted. By using two consecutive 3x3 convolutional layers and Leaky ReLU activation functions in the Inception module, and adding data standardized operations after each module, the network structure is optimized to improve recognition performance.
It improves the accuracy and generalization ability of public facilities identification, reduces the computational complexity and parameter quantity, enhances the nonlinear processing capability and stability of the network, and significantly improves the identification efficiency.
Smart Images
Figure CN119516331A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of public facility detection, and in particular to a public facility detection method based on a convolutional neural network GoogleNet. Background Art
[0002] With the development of cities, the various components of cities interact with each other in different forms and ways to form various urban structures. Among them, the public facilities in the blocks help us improve the quality of life. However, with the increase in facilities, the increase in service life, and the frequent use of facilities, the current problems of damage and aging of facilities have emerged, and the inspection and maintenance of public facilities are also particularly important. At present, the manual inspection method of public facilities is time-consuming and labor-intensive, and can no longer meet the current needs. In the future, artificial intelligence will inevitably be used in the intelligent management of blocks.
[0003] CNN has shown extraordinary performance in many fields, and many scholars have begun to try to apply CNN to the field of image recognition. CNN does not require additional feature extraction when performing classification and recognition. The entire feature extraction and classification process is implemented in one link, so the model structure is simple and the implementation process is relatively easy, which can solve the image recognition problem well.
[0004] GoogLeNet is a 22-layer deep network that uses a unique Inception structure to reduce network parameters. General CNNs only use the method of increasing the network to improve network performance, which has the risk of overfitting during training and increases the computational cost. The Inception structure solves these two problems well. It retains the sparsity of the network structure and the high computational performance of the dense matrix, and reduces the parameters while increasing the width and depth of the network.
[0005] Nowadays, the field of image recognition is still in its infancy in the field of facility recognition and detection, but it is widely used in some other object recognition fields. The entire feature extraction and classification process is implemented in one link, which can solve the image recognition problem well, but there are also some problems. The current recognition methods not only have low recognition accuracy, but also require additional features to be applied, and the structure is complex. Traditional image recognition methods have the risk of overfitting, low accuracy, and insufficient generalization ability. Summary of the invention
[0006] The technical solution of the present invention to solve the above technical problems is to provide a public facility detection method based on a convolutional neural network GoogleNet, comprising the following steps:
[0007] Step 1: Collect public facility image datasets and preprocess them; the preprocessed image datasets are divided into training sets and test sets;
[0008] Step 2: Build the GoogLeNet network model. The Inception module includes two consecutive 3x3 convolutional layers and a Leaky ReLU activation function. The output of each 3x3 convolutional layer is processed by a 3x3 convolution kernel until the final feature map is generated. The convolution method of group convolution is adopted.
[0009] Add data normalization operation after each Inception module to correct the distribution of input data and prevent gradient disappearance; determine the optimal bias learning rate factor of the fully connected layer, set the initial learning rate to 0.0001, and the learning rate factor to 10;
[0010] The standardization operation is:
[0011]
[0012] where μ is the mean, σ 2 is the variance, x i is the feature map in the batch and m is the batch size.
[0013]
[0014] Here, ε is a small constant used to prevent division by zero errors.
[0015]
[0016] Where γ and β are learnable parameters for scaling and translation, respectively.
[0017] Step 3: Input the preprocessed test set into the GoogLeNet network model to complete the identification and detection of public facilities.
[0018] Furthermore, in step 1, the preprocessing includes randomly rotating the image with a rotation angle ranging from -60° to 60°, scaling the image to 0.8 to 1.2 times of the original size, and randomly translating the image in the horizontal and vertical directions with a translation pixel range of -30 to 30.
[0019] Furthermore, in step 1, the public facility image dataset is obtained through field collection, including five public facility categories: signboards, public trash cans, street lights, fire hydrants, and public seats.
[0020] Furthermore, in step 1, the training set accounts for 80% of the image data set, and the test set accounts for 20% of the image data set.
[0021] Furthermore, in step 2, the Leaky ReLU activation function assigns a smaller slope to negative input values, usually with a value between 0.01 and 0.1.
[0022] Compared with the prior art, this application has the following beneficial effects:
[0023] 1. Use two consecutive 3x3 convolutional layers to replace the original single 5x5 convolutional layer to optimize the Inception module, and use group convolution instead of normal convolution kernels to improve the computational efficiency of public facility recognition by improving the original Googlenet network.
[0024] 2. The Leaky ReLU activation function is introduced to replace the original ReLU activation function, which enhances the nonlinear processing ability of the network and thus improves the generalization. It has a positive effect on the performance in public facility identification.
[0025] 3. Data standardization is added after each Inception module, so that the model can stably learn useful regional features during training, thereby more reliably extracting the features of public facilities. The fully connected layer further accelerates the convergence of the network by setting a specific bias learning rate factor. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.
[0027] Figure 1 It is a structural schematic diagram of the public facilities detection method based on the convolutional neural network GoogleNet of the present invention;
[0028] Figure 2 It is a schematic diagram of the structure of the group convolution of the present invention;
[0029] Figure 3 It is a structural diagram of the Inception module of the present invention;
[0030] Figure 4 This is a comparison chart of the ReLU and Leaky ReLU activation function curves of the present invention;
[0031] Figure 5 This is a schematic diagram of the Googlenet network model structure of the present invention;
[0032] Figure 6This is a comparison chart of the verification set accuracy of the original GoogLeNet network and the improved GoogLeNet network of the present invention;
[0033] Figure 7 It is a comparison diagram of the loss value curves of the original GoogLeNet network and the improved GoogLeNet network of the present invention;
[0034] Figure 8 This is a diagram of the public facilities identification results of the present invention. DETAILED DESCRIPTION
[0035] The present invention proposes a public facility detection method based on a convolutional neural network GoogleNet, aiming to design a public facility detection method with high accuracy and strong generalization ability.
[0036] The public facilities detection method based on the convolutional neural network GoogleNet proposed by the present invention will be described below in a specific embodiment:
[0037] In the technical solution of this embodiment, Figure 1 As shown, a public facility detection method based on a convolutional neural network GoogleNet includes the following steps:
[0038] Step 1: Collect public facility image datasets and preprocess them; the preprocessed image datasets are divided into training sets and test sets;
[0039] Step 2: Build the GoogLeNet network model. The Inception module includes two consecutive 3x3 convolutional layers and a Leaky ReLU activation function. The output of each 3x3 convolutional layer is processed by a 3x3 convolution kernel until the final feature map is generated. The convolution method of group convolution is adopted.
[0040] Add data normalization operation after each Inception module to correct the distribution of input data and prevent gradient disappearance; determine the optimal bias learning rate factor of the fully connected layer, set the initial learning rate to 0.0001, and the learning rate factor to 10;
[0041] The standardization operation is:
[0042]
[0043]
[0044] where μ is the mean, σ 2 is the variance, x i is the feature map in the batch and m is the batch size.
[0045]
[0046] Here, ε is a small constant used to prevent division by zero errors.
[0047]
[0048] Where γ and β are learnable parameters for scaling and translation, respectively.
[0049] Specifically, a data standardization operation is added after each Inception module. As the number of network layers increases, the distribution of input data in each layer also changes. The addition of data standardization can correct the distribution of input data and standardize the input data. The data distribution can be adjusted to around 0 to prevent gradient disappearance. Through experiments, the fully connected layer obtains the optimal bias learning rate factor, which further accelerates the convergence speed of the network. The data that needs to be standardized is X = {x1, x2, ..., x n}, n is the size of the data, and the result after the standardization operation is Y = {y1, y2, ..., y m}.
[0050] Step 3: Input the preprocessed test set into the GoogLeNet network model to complete the identification and detection of public facilities.
[0051] Furthermore, in step 1, the preprocessing includes randomly rotating the image with a rotation angle ranging from -60° to 60°, scaling the image to 0.8 to 1.2 times of the original size, and randomly translating the image in the horizontal and vertical directions with a translation pixel range of -30 to 30.
[0052] Furthermore, in step 1, the public facility image dataset is obtained through field collection, including five public facility categories: signboards, public trash cans, street lights, fire hydrants, and public seats.
[0053] Furthermore, in step 1, the training set accounts for 80% of the image data set, and the test set accounts for 20% of the image data set.
[0054] Furthermore, in step 2, the Leaky ReLU activation function assigns a smaller slope to negative input values, usually with a value between 0.01 and 0.1.
[0055] like Figure 2 , Figure 3 , Figure 4 , Figure 5As shown, this application aims to solve the problems of high computational complexity and large number of parameters in the Inception module by using two consecutive 3x3 convolutional layers to replace the original single 5x5 convolutional layer to optimize the Inception module, and using group convolution instead of normal convolution kernels, which improves the feature extraction capability and network flexibility, and has better real-time processing capabilities; for the processing of complex data and to avoid "dead neurons", the Leaky ReLU activation function is introduced to replace the original ReLU activation function, which enhances the nonlinear processing capability of the network; after each Inception module, a data standardization operation is added, which can correct the distribution of the input data and standardize the input data to prevent the gradient from disappearing; finally, the fully connected layer obtains the optimal bias learning rate factor through experiments, and sets the learning rate factor to 10. Set the initial learning rate to 0.0001.
[0056] The evaluation index uses recognition accuracy to evaluate the classification performance of the model, which represents the ratio of correctly predicted samples to the total predicted samples:
[0057]
[0058] Where: TP represents true positive examples; TN represents true negative examples; FP represents false positive examples; FN represents false negative examples
[0059] This application selects AlexNet, VGG16, ResNet 18 to compare with the GoogLeNet network model optimized by this application. Table 1 shows the accuracy and error rate details of each model. The optimized GoogLeNet can stabilize faster than other models. In terms of training accuracy, the improved GoogleNet has the highest accuracy of 98.15%, which is higher than ResNet18, VGG16 and AlexNet.
[0060] Table 1: Comparative experiment:
[0061]
[0062] like Figure 6 As shown in Figure 2, this application monitors the impact of the Leaky Relu activation function and the improved Inception on GoogLeNet by controlling variables. In order to compare the performance of the Leaky Relu and ReLU functions, we introduce these two functions into GoogLeNet respectively, and the comparison results are shown in Table 2. The Leaky Relu function improves the accuracy by 0.13% over the ReLU function.
[0063] Table 2
[0064]
[0065] In order to understand the performance of Leaky Relu, a comparative experiment was conducted and ReLu was selected as the activation function. From Table 3, we can see that the improved Inception effect is relatively good, and the accuracy is improved by 0.42%.
[0066] Table 3
[0067]
[0068] Finally, Leaky Relu and Leaky Relu are introduced into GoogLeNet at the same time, and the following data are obtained, as shown in Table 4. It is found that the training accuracy of GoogLeNet is improved by 0.95%. This shows that the two experiments can improve the training accuracy of GoogLeNet.
[0069] Table 4
[0070]
[0071] The model was trained on the same image dataset to ensure that each model was trained in the same experimental environment. During the training process, the accuracy of the model was verified and the data was saved after each training cycle. After the introduction of the attention mechanism, the recognition accuracy of the model was greatly improved, but the size of the model increased very little, further demonstrating the superiority of the model performance. The accuracy of the original Googlenet network on the verification dataset and the memory capacity occupied by the model are compared in Table 5.
[0072] Table 5
[0073]
[0074] This application compares GoogLeNet with the GoogLeNet optimized by this application and finds that the training accuracy of the optimized model is 98.15%, which is 1.32% higher than GoogLeNet. This training accuracy is much higher than the two attempts made to improve the training accuracy. In order to reduce the number of neurons and ensure the comprehensiveness of feature extraction, full dense connections and the help of two convolution kernels of different sizes are used. This design makes our model feature extraction more powerful, and perhaps this is also the reason why our model training accuracy has been greatly improved. Figure 7 As shown in the figure, the loss value of the improved Googlenet network model converges faster and has better performance than the original Googlenet network model.
[0075] Finally, the weights obtained by training the optimized Googlenet network model are used to predict the randomly selected images, and the recognition results are shown in the figure below. Figure 8As shown in the figure, the prediction results all achieved an accuracy rate of more than 98%.
[0076] Compared with the original GoogLeNet:
[0077] 1. The original GoogLeNet processing efficiency is insufficient. The Inception module in the original GoogLeNet network is optimized by using two consecutive 3x3 convolutional layers instead of the original single 5x5 convolutional layer, and group convolution is used instead of normal convolution kernels. This change not only reduces the number of network parameters and computational complexity, but also improves the computational efficiency of public facility recognition. This is very important for tasks such as public facility recognition that may have multiple backgrounds and environmental changes.
[0078] 2. The original GoogLeNet has insufficient generalization ability. When the input enters the ReLU activation function, if the result is a negative number, the gradient will become zero, causing the neuron to no longer update the weight during training, causing some neurons in the network to "die" and can no longer be activated. The introduction of the Leaky ReLU activation function to replace the original ReLU activation function enhances the nonlinear processing ability of the network and has a positive effect on the performance in public facility recognition.
[0079] 3. The original GoogLeNet may not have sufficient feature processing capabilities in the recurrent context. The addition of data standardization after each Inception module, complex block backgrounds and diverse public facility data sets may lead to unstable training processes. Through standardization, the input feature distribution is more consistent, allowing the model to stably learn useful regional features during training, thereby more reliably extracting features of public facilities and preventing gradient disappearance. Finally, the fully connected layer further accelerates the convergence of the network by setting a specific bias learning rate factor.
[0080] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A public facility detection method based on convolutional neural network GoogleNet, characterized in that: The following steps are involved: Step 1: Collect public facility image datasets and preprocess them; the preprocessed image datasets are divided into training sets and test sets; Step 2: Build the GoogLeNet network model. The Inception module includes two consecutive 3x3 convolutional layers and LeakyReLU activation function. The output of each 3x3 convolutional layer is processed by a 3x3 convolution kernel until the final feature map is generated. The convolution method of group convolution is adopted. Add data normalization operation after each Inception module to correct the distribution of input data and prevent gradient disappearance; determine the optimal bias learning rate factor of the fully connected layer, set the initial learning rate to 0.0001, and the learning rate factor to 10; The standardization operation is: where μ is the mean, σ 2 is the variance, x i is the feature map in the batch, m is the batch size; Among them, ε is a small constant used to prevent division by zero errors; Where γ and β are learnable parameters, used for scaling and translation respectively; Step 3: Input the preprocessed test set into the GoogLeNet network model to complete the identification and detection of public facilities.
2. The public facility detection method according to claim 1, characterized in that: In step 1, the preprocessing includes randomly rotating the image with a rotation angle ranging from -60° to 60°, scaling the image to 0.8 to 1.2 times of the original size, and randomly translating the image in the horizontal and vertical directions with a translation pixel range of -30 to 30.
3. The public facility detection method according to claim 1, characterized in that: In step 1, the public facility image dataset is obtained through field collection, including five public facility categories: signboards, public trash cans, street lights, fire hydrants, and public seats.
4. The public facility detection method according to claim 1, characterized in that: In step 1, the training set accounts for 80% of the image data set, and the test set accounts for 20% of the image data set.
5. The public facility detection method according to claim 1, characterized in that: In step 2, the Leaky ReLU activation function assigns a slope of 0.01 to 0.1 to negative input values.