A Traffic Sign Classification Method Based on Deep Convolutional Fuzzy System
By proposing a traffic sign classification method based on a deep convolutional fuzzy system, this method utilizes convolution and pooling operations to extract features, combines PCA dimensionality reduction and one-hot encoding transformation, and uses DFSSR for decision-making. This solves the problems of model complexity and poor interpretability in traffic sign recognition of deep convolutional neural networks, achieving high-precision and efficient classification results.
Patent Information
- Application Number
- CN202411939181.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-26
- Publication Date
- 2025-12-02
- Estimated Expiration
- 2044-12-26
AI Technical Summary
Existing deep convolutional neural networks suffer from problems such as model complexity, numerous hyperparameters, and poor interpretability in traffic sign recognition, especially in extreme weather conditions and blurry image conditions where they are difficult to classify effectively.
A traffic sign classification method based on a deep convolutional fuzzy system is adopted. Features are extracted through convolution and pooling operations, PCA is used for dimensionality reduction, one-hot encoding is used to transform the multi-class problem into multiple binary classification problems, and DFSSR is used to solve them. Finally, the output is integrated through the softmax function.
It improves the accuracy and interpretability of traffic sign classification, simplifies the model structure, reduces the number of parameters, avoids the gradient vanishing problem, and improves training efficiency and performance.
Smart Images
Figure CN119851192B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of traffic sign recognition and classification technology, and in particular to a traffic sign classification method based on a deep convolutional fuzzy system. Background Technology
[0002] The development of Convolutional Neural Networks (CNNs) has opened up new avenues for traffic sign recognition research, further advancing the study of traffic sign recognition and classification. Mehta et al. proposed a traffic sign classification method based on deep convolutional networks, employing Adam (Adaptive Moment Estimation) as the optimizer and softmax as the activation in the output layer. Tests on the Belgian Traffic Signs Dataset (BTSD) showed that this traffic sign classification method is highly competitive compared to previous algorithms. Furthermore, a novel CNN architecture has been proposed to improve traffic sign detection performance. This method uses the Adam optimizer, sets the batch size to 128, and employs multiple interconnected layers, enabling it to detect traffic signs even under extreme weather conditions and blurred image conditions.
[0003] While deep convolutional neural networks (CNNs) possess excellent learning capabilities and high prediction accuracy, they still suffer from problems such as model complexity, numerous hyperparameters, and black-box nature. As the scale of data continues to expand, the connections between neurons in deep neural networks become increasingly complex, ultimately making the entire model difficult to understand. Therefore, improving deep learning algorithms and enhancing the interpretability of traffic sign classification methods has become a pressing scientific problem. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a traffic sign classification method based on a deep convolutional fuzzy system.
[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:
[0006] A traffic sign classification method based on a deep convolutional fuzzy system includes:
[0007] 1) Extract features from traffic sign image data using convolution and pooling operations;
[0008] 2) Use PCA to reduce the dimensionality of the feature data extracted from traffic sign images;
[0009] 3) Using one-hot encoding, the multi-classification problem is transformed into multiple binary classification problems for solving. Each binary classification problem is solved using DFSSR (Deep Fuzzy System with Simplified Rule base).
[0010] 4) Integrate the outputs of all DFSSR systems, process them using the softmax function, and obtain the output of the DCFSSR (Deep Convolutional Fuzzy System with Simplified Rule base) system.
[0011] As one possible implementation, further, in step 1), ResNet18 is used as a feature extractor to convert traffic sign image data into feature data.
[0012] As one possible implementation, the ResNet18 further includes:
[0013] Input layer: Used to receive image data;
[0014] Convolutional layer: The input image is processed using a 7×7 convolutional kernel with a stride of 2, and then processed through a max pooling layer with the same stride of 2 to reduce the spatial dimension of the feature map and increase the abstraction of the features.
[0015] Residual blocks: Each residual block contains two 3×3 convolutional layers, including activation functions and batch normalization operations. The input of each block is not only passed to the next convolutional layer, but also directly added to the block output through skip connections, which promotes the direct flow of gradients.
[0016] Global average pooling layer: After all residual blocks, the global average pooling layer calculates the average value of each feature map channel, converting the multi-dimensional feature map into a one-dimensional feature vector, providing input for the fully connected layer;
[0017] Fully connected layer: connects the feature vectors after global average pooling to the final classification layer;
[0018] Softmax layer: Calculates the final output, which is the predicted probability distribution for each category.
[0019] As one possible implementation, further, in step 2), feature data with a cumulative contribution of 95% are selected, and PCA is used to reduce the dimensionality of the data.
[0020] As one possible implementation, the DFSSR further comprises several sub-modules FSSR (FuzzySystem with Simplified Rule base), with FSSRs stacked layer by layer; the input data for each FSSR is set to k-dimensional, and each FSSR module uses at most r rules; each binary classifier DFSSR outer structure has a total of Layer, in which This indicates rounding up, where N is the dimension of the data features after PCA dimensionality reduction.
[0021] As one possible implementation, the FSSR algorithm further includes the following steps:
[0022] Divide the data into fuzzy sets and define the corresponding membership functions;
[0023] Rules are learned from the data based on the defined fuzzy sets, with a one-to-one correspondence between the data and the rules. At the same time, the contribution of each rule is calculated during the rule learning process.
[0024] After the rules have been learned, rules with the same antecedents are merged, and the importance of each merged rule is calculated.
[0025] The rules are sorted according to their importance, and the key rules are extracted to form the rule base of FSSR.
[0026] As one possible implementation, the DCFSSR further comprises C DFSSRs, and the C DFSSRs contain There are FSSRs, where C represents the number of classes in the classification problem;
[0027] During the construction of the binary classifier DFSSR, for the same DCFSSR, its binary classifier DFSSRs all use the same grouping and block scheme, and each DFSSR uses the same system input and shares the same rule base in the first level.
[0028] As one possible implementation, the training method for the binary classifier DFSSR in step 3) further includes the following steps:
[0029] S1: Determine k L Check if =N is true; if true, proceed to the next step; otherwise, start from the system's input data (x1, x2, ..., x...). N Randomly select (k) from ) L -N) feature data, reuse feature data, and pad the data to k. L ;
[0030] S2: Randomly group the feature data, with each group consisting of k data features. During the grouping process, it must be ensured that no two data features are the same in each group; otherwise, the groups are regrouped.
[0031] S3: Input the divided grouped data into each sub-module FSSR for learning. Sub-module FSSRs in the same level do not interfere with each other and learn in parallel.
[0032] S4: Integrate the output of the FSSR submodule and input it to the next layer, where the FSSR submodule in the next layer continues to learn;
[0033] S5: Regroup the data according to the input order, with each group consisting of k data features; if the system has k input data at this time, execute S6; otherwise, continue from S3.
[0034] S6: Integrate these k feature data and input them into the L-layer (i.e., top-layer) submodule FSSR. The final decision is made by the last-layer FSSR, resulting in the system output y' of the entire DFSSR.
[0035] As one possible implementation, step 4) further includes the following steps:
[0036] Integrate the outputs z1, z2, ..., z of all DFSSR systems C Where z1, z2, ..., z C This represents the classification results corresponding to categories 1 through C of the DCFSSR.
[0037] After calculation using the following formula (4), the result probability of each binary classifier DFSSR can be obtained;
[0038]
[0039] in, 0 < y j <1, j = 1, 2, ..., C.
[0040] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:
[0041] 1) This invention constructs a deep fuzzy system, DFSSR, which first decomposes the complex high-dimensional problem into multiple sub-problems, employing a divide-and-conquer strategy. Then, the FSSR is stacked layer by layer to integrate the sub-problems and solve the overall problem. Compared to other algorithms, the DCFSSR provided by this invention exhibits higher accuracy. Compared to the ResNet18 model, DCFSSR completes classification decisions through rules, possessing not only better interpretability but also comparable prediction accuracy to ResNet18, while requiring fewer parameters; therefore, DCFSSR has a greater advantage in traffic sign classification problems.
[0042] 2) After extracting image features, DCFSSR relies on a fuzzy system for subsequent decision-making processes. In this way, DCFSSR can complete the decision-making process for traffic sign classification through IF-THEN rules, thereby enhancing the interpretability of the decision-making process.
[0043] 3) The ResNet18 structure in this invention effectively avoids the vanishing gradient problem, enabling the network to be trained effectively even with increased depth. Furthermore, the residual block design reduces the training difficulty of the model, allowing a portion of the network layers to focus on learning the residual portion of the input data, rather than directly learning the entire feature map from scratch. This simplifies the learning process, thereby improving the network's training efficiency and final performance. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a simplified flowchart of the present invention;
[0046] Figure 2 This is a schematic diagram of the internal structure of the FSSR;
[0047] Figure 3 This is a schematic diagram of the DFSSR topology. Detailed Implementation
[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] This invention, based on research into the Deep fuzzy system DFSSR and incorporating the achievements of deep convolutional neural networks in image feature extraction, proposes the Deep Convolutional Fuzzy System Algorithm DCFSSR by fusing one-hot encoding and the softmax function. DCFSSR uses convolution and pooling techniques to extract image features, and then performs dimensionality reduction on the feature data using PCA. Next, one-hot encoding is used to decompose the multi-class problem into several binary classification problems, each of which is solved by the DFSSR algorithm. Finally, all outputs of the binary classifier DFSSR are integrated and processed using the softmax function to obtain the predicted class labels from the DCFSSR system.
[0050] See attached document Figure 1 As shown, this embodiment provides a traffic sign classification method based on a deep convolutional fuzzy system, including:
[0051] 1) Features are extracted from traffic sign image data through convolution and pooling operations. In this embodiment, ResNet18 is used as a feature extractor to transform traffic sign image data into feature data.
[0052] ResNet18 includes:
[0053] Input layer: Used to receive image data.
[0054] Convolutional layer: First, a 7×7 convolutional kernel with a stride of 2 is used to process the input image. Then, a max pooling layer with the same stride of 2 is used to reduce the spatial dimension of the feature map and increase the abstraction of the features.
[0055] Residual blocks: The core of ResNet. Each residual block contains two 3×3 convolutional layers, including activation functions and batch normalization operations; the input of each block is not only passed to the next convolutional layer, but also directly added to the block output through skip connections, promoting direct gradient flow.
[0056] Global average pooling layer: After all residual blocks, the global average pooling layer calculates the average value of each feature map channel, transforming the multidimensional feature map into a one-dimensional feature vector, providing input for the fully connected layer.
[0057] Fully connected layer: Connects the feature vectors after global average pooling to the final classification layer.
[0058] Softmax layer: Calculates the final output, which is the predicted probability distribution for each category.
[0059] 2) Use PCA (Principal Component Analysis) to reduce the dimensionality of the feature data extracted from the traffic sign images; specifically, select the feature data with a cumulative contribution of 95%, use PCA to reduce the dimensionality of the data, and provide it to DCFSSR for multi-classification tasks.
[0060] 3) Using one-hot encoding, the multi-class classification problem is transformed into multiple binary classification problems for solution. Each binary classification problem is solved using DFSSR (as shown in the appendix). Figure 3 (As shown); where DFSSR contains several sub-modules FSSR, and FSSRs are stacked layer by layer;
[0061] See attached document Figure 2As shown, the algorithm steps of FSSR (Fuzzy System with Simplified Rule base) are as follows: First, the data is divided into fuzzy sets, and corresponding membership functions are defined. Then, rules are learned from the data based on the defined fuzzy sets, one rule per data point. During the rule learning process, the contribution of each rule is calculated. After rule learning is complete, rules with the same antecedents are merged. After merging, the importance of each merged rule is calculated. Then, the rules are ranked according to their importance, and the key rules are extracted as the rule base for FSSR. Specifically, the FSSR algorithm is shown in Table 1 below:
[0062] Table 1 FSSR Algorithm
[0063]
[0064]
[0065] The formulas in the table above are shown below:
[0066]
[0067] In the above formula, N represents the feature dimension of the data. Represents the fuzzy operator index. Let r represent the membership value of the r-th rule in the N-th dimension of the fuzzy set;
[0068]
[0069] In the above formula, This represents the membership value of the r-th rule in the N-th dimension of the fuzzy set. Represents the fuzzy operator index;
[0070]
[0071] In the above formula, Let y1 represent the nth dimension of the m-th data point, where y1...y2...y3... m This represents the system's predicted output. This represents the number of rules after filtering. This indicates that the Nth dimension of the Mth data point is in the... The trigger strength obtained from the rule, where M represents the number of samples and N represents the dimension. This means that the parameter series ω is calculated using the least squares method in conjunction with formula (2-21).
[0072] The FSSR system input is set to k (the input data of each FSSR is k-dimensional), and the outer structure of each binary classifier DFSSR has a total of Layer, in which This indicates rounding up, where N is the dimension of the data features after PCA dimensionality reduction.
[0073] All binary classifiers, such as DFSSR, need to be trained. The training method for DFSSR includes the following steps:
[0074] S1: Determine k L Check if =N is true; if true, proceed to the next step; otherwise, start from the system's input data (x1, x2, ..., x...). N Randomly select (k) from ) L -N) feature data, reuse feature data, and pad the data to k. L ;
[0075] S2: Randomly group the feature data, with each group consisting of k data features. During the grouping process, it must be ensured that no two data features are the same in each group; otherwise, the groups are regrouped.
[0076] S3: Input the divided grouped data into each sub-module FSSR for learning. Sub-module FSSRs in the same level do not interfere with each other and learn in parallel.
[0077] S4: Integrate the output of the FSSR submodule and input it to the next layer, where the FSSR submodule in the next layer continues to learn;
[0078] S5: Regroup the data according to the input order, with each group consisting of k data features; if the system has k input data at this time, execute S6; otherwise, continue from S3.
[0079] S6: Integrate these k feature data and input them into the L-layer (i.e., top-layer) submodule FSSR. The final decision is made by the last-layer FSSR, resulting in the system output y' of the entire DFSSR.
[0080] For the submodules in the first layer of DFSSR, their system output can be represented as:
[0081]
[0082] Where w1={z1,z2,...,z k}, w2={z k+1 ,z k+2 ,...,z 2k}, R1, R2, Rk represent the system inputs of the 1st, 2nd, and kth modules in the first layer of DFSSR, respectively. k This represents the number of rules learned. And b r,n ,c r,n ,d r,nLet be the consequent parameter of the r-th rule in the N-th dimension obtained after the system training is completed. This represents the normalized rule trigger strength. Data x = (x1, x2, ..., x...). N After feature reuse and random sorting, a new data sequence was obtained. Among them, z i ∈{x1,x2,...,x N}, i = 1, 2, ..., k 2 N is the feature dimension of the input data.
[0083] The DFSSR proposed in this invention utilizes random feature reuse to ensure that all FSSR modules have the same input size, thereby reducing the differences between FSSRs. This allows each FSSR module to use a uniform parameter configuration, thus maintaining consistency among FSSR modules. Each FSSR module has five internal layers, and the DFSSR has an outer layer of L layers, resulting in a total of 5L layers in the entire DFSSR.
[0084] 4) After training, integrate the output data of all DFSSR systems, process them using the softmax function, and obtain the output of the DCFSSR system (where DCFSSR outputs the probabilities corresponding to various classification results); specifically including the following steps:
[0085] Integrate the outputs z1, z2, ..., z of all DFSSR systems C Where z1, z2, ..., z C This represents the classification results for classes 1 through C of the DCFSSR, where C represents the total number of classes in the classification problem.
[0086] After calculation using the following formula (4), the result probability of each binary classifier DFSSR can be obtained;
[0087]
[0088] in, 0 < y j <1, j = 1, 2, ..., C.
[0089] In the construction of the binary classifier DFSSR, all binary classifier DFSSRs adopt the same deep structure construction method; that is, for the same DCFSSR, its binary classifier DFSSRs use the same grouping and block partitioning scheme. The input data of the FSSR module is fixed at k dimensions, and the outer structure of each binary classifier DFSSR has a total of... Therefore, a DCFSSR needs to construct C DFSSRs. Each module FSSR.
[0090] Since each binary classifier (DFSSR) in DCFSSR uses the same grouping and block partitioning pattern and the same system input for the first level, each DFSSR shares the same rule base for its first level; that is, each DFSSR uses the same rule base for its first level, the difference being in the consequent parameters. Therefore, assuming each module FSSR uses r rules, the overall construction of DCFSSR requires... There are 10 rules in total. One parameter.
[0091] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A traffic sign classification method based on a deep convolutional fuzzy system, characterized in that, include: 1) Extract features from traffic sign image data using convolution and pooling operations; 2) Use PCA to reduce the dimensionality of the feature data extracted from traffic sign images; 3) Using one-hot encoding, the multi-classification problem is transformed into multiple binary classification problems for solution. Each binary classification problem is solved using DFSSR. The DFSSR contains several sub-modules FSSR, which are stacked layer by layer. The input data for each FSSR is set to k dimensions, and each FSSR module uses at most r rules. Each binary classifier DFSSR has a total of [number missing] outer layers. Layer, in which This indicates rounding up, where N is the dimension of the data features after PCA dimensionality reduction; 4) Integrate the outputs of all DFSSR systems, process them using the softmax function, and obtain the output of the DCFSSR system; the DCFSSR contains C DFSSRs, and the C DFSSRs contain There are FSSRs, where C represents the number of classes in the classification problem; During the construction of the binary classifier DFSSR, for the same DCFSSR, its binary classifier DFSSRs all use the same grouping and block scheme, and each DFSSR uses the same system input and shares the same rule base in the first level.
2. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 1, characterized in that, In step 1), ResNet18 is used as the feature extractor to convert traffic sign image data into feature data.
3. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 2, characterized in that, The ResNet18 includes: Input layer: Used to receive image data; Convolutional layer: The input image is processed using a 7×7 convolutional kernel with a stride of 2, and then processed through a max pooling layer with the same stride of 2. Residual blocks: Each residual block contains two 3×3 convolutional layers, including activation functions and batch normalization operations. The input of each block is not only passed to the next convolutional layer, but also directly added to the block output through skip connections, which promotes the direct flow of gradients. Global average pooling layer: After all residual blocks, the global average pooling layer calculates the average value of each feature map channel, converting the multi-dimensional feature map into a one-dimensional feature vector, providing input for the fully connected layer; Fully connected layer: connects the feature vectors after global average pooling to the final classification layer; Softmax layer: Calculates the final output, which is the predicted probability distribution for each category.
4. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 1, characterized in that, Step 2) Select the feature data with a cumulative contribution of 95% and use PCA to reduce the dimensionality of the data.
5. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 1, characterized in that, The FSSR algorithm steps include: Divide the data into fuzzy sets and define the corresponding membership functions; Rules are learned from the data based on the defined fuzzy sets, with a one-to-one correspondence between the data and the rules. At the same time, the contribution of each rule is calculated during the rule learning process. After the rules have been learned, rules with the same antecedents are merged, and the importance of each merged rule is calculated. The rules are sorted according to their importance, and the key rules are extracted to form the rule base of FSSR.
6. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 1, characterized in that, The training method for the binary classifier DFSSR described in step 3) includes the following steps: S1: Judgment Check if the condition is met; if so, proceed to the next step; otherwise, check the system's input data. Randomly selected from ( ) feature data, reuse feature data, and fill in the data to complete the set. ; S2: Randomly group the feature data, each... k Each data feature is divided into a group. During the division process, it must be ensured that no two data features are the same in each group; otherwise, the groups are re-divided. S3: Input the divided grouped data into each sub-module FSSR for learning. Sub-module FSSRs in the same level do not interfere with each other and learn in parallel. S4: Integrate the output of the FSSR submodule and input it to the next layer, where the FSSR submodule in the next layer continues to learn; S5: Group the data again according to the input order, each... k Each data feature is grouped into a set; if the system's input data at this time is k If the result is positive, then execute S6; otherwise, continue executing from S3. S6: Integrate this k Each feature data point is input into... L In the FSSR submodule of the layer, the final decision is made by the FSSR of the last layer, resulting in the system output of the entire DFSSR. y’ .
7. The traffic sign classification method based on a deep convolutional fuzzy system according to claim 1, characterized in that, Step 4) specifically includes the following steps: Integrate the outputs of all DFSSR systems ,in This represents the classification results corresponding to categories 1 through C of the DCFSSR. After calculation using the following formula (4), the result probability of each binary classifier DFSSR can be obtained; （4） in, , , .