An improved capsule neural network method for handwritten signature recognition
By adopting adaptive learning clustering algorithm and Cronek product method in capsule neural network, combined with least squares support vector machine, the problem of overfitting and slow recognition speed in handwritten signature recognition is solved, achieving more efficient feature extraction and recognition accuracy.
Patent Information
- Application Number
- CN202110233329.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-03-03
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-03-03
AI Technical Summary
Existing capsule neural networks are prone to overfitting problems and slow recognition speed in handwritten signature recognition, and require a large number of samples for training.
Adaptive learning clustering algorithm is used for feature extraction, Cronek product method is used for parameter calculation, and a least squares support vector machine is built as a classifier to improve the capsule neural network structure, reduce the amount of parameters and calculation, and improve the recognition efficiency .
It effectively avoids overfitting problems, reduces sample size requirements, improves recognition speed and accuracy, and builds a more compact deep neural network model.
Smart Images

Figure CN113011289B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of computer vision and artificial intelligence, and in particular relates to a handwritten signature recognition method of an improved capsule neural network. Background Art
[0002] The handwritten signature recognition system is a hot topic in text classification. The handwritten signature recognition system mainly involves knowledge in image processing and pattern recognition, including image preprocessing, information area positioning and recognition, and has been widely used in information processing, identity authentication, and human-computer interaction. However, in real life, handwritten signatures may have different degrees of differences due to the psychological state of the signer and the environment in which they are located. Therefore, signature feature extraction and feature representation are very complicated, so the recognition accuracy is not high. Improving the recognition rate of handwritten signatures has become an important research hotspot.
[0003] The use of machine learning methods for image recognition mainly depends on the extraction of image features. The effectiveness of feature extraction affects the accuracy of recognition. As a new branch of machine learning, deep learning analyzes and processes data information by simulating the human brain mechanism. Capsule neural networks in deep learning have made major breakthroughs in the field of image recognition. They can self-learn the features of images from low-level to high-level, avoiding the shortcomings of manual feature extraction in machine learning. However, capsule neural networks require the support of massive samples during training, and when there is insufficient data, overfitting problems will occur.
[0004] Therefore, a new technical solution is needed to solve the overfitting problem and slow recognition speed problem that occur when using capsule neural networks for image recognition. Summary of the invention
[0005] The present invention provides an improved capsule neural network handwritten signature recognition method, which solves the overfitting problem and slow recognition speed problem that may occur when using traditional capsule neural networks for image recognition in the prior art.
[0006] In order to achieve the above object, the technical solution of the present invention is:
[0007] An improved capsule neural network handwritten signature recognition method comprises the following steps:
[0008] Step 1: Obtain and preprocess the handwritten signature dataset;
[0009] Step 2, building a basic model of the improved capsule neural network, in which an adaptive learning clustering algorithm is used for feature extraction and a Kronecker product method is used for parameter calculation;
[0010] Step 3: Build a least squares support vector machine as a classifier;
[0011] Step 4, training: use the preprocessed handwritten signature dataset as the sample library to train the improved capsule neural network and set several epochs until the training accuracy is stable;
[0012] Step 5: Input the features extracted by the capsule neural network into the least squares support vector machine to train the least squares support vector machine;
[0013] Step 6: Prediction: Make predictions on new unknown images based on the trained handwritten signature recognition model.
[0014] Furthermore, in step (2), the collected image is first preprocessed, including grayscale processing, tilt processing, and denoising processing, and then the handwritten signature is cut and extracted as the information area to be identified to obtain an image with only the handwritten signature.
[0015] Compared with the prior art, the present invention has the following beneficial effects:
[0016] The present invention proposes an improved capsule neural network handwritten signature recognition method, which uses the capsule neural network method to avoid the shortcomings of manually extracting features, solves the problem of deep learning taking a long time and consuming resources by improving the capsule neural network method, and improves the disadvantage of capsule neural network requiring a large number of samples to support training by using the least squares support vector machine as a classifier. The deep neural network model established by this method has a more compact structure, requires a smaller sample size, and greatly reduces the number of parameters generated. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is the overall scheme diagram of the present invention;
[0018] Figure 2 It is a schematic diagram of the adaptive clustering algorithm;
[0019] Figure 3 Schematic diagram of replacing the fully connected layer with the Kronecker product;
[0020] Figure 4 It is a schematic diagram of the overall process of the present invention;
[0021] Figure 5 The following is a comparison chart of the experimental results;
[0022] Figure 6 This is a comparison curve of the recognition accuracy of the method of the present invention and the traditional capsule neural network. DETAILED DESCRIPTION
[0023] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0024] Example:
[0025] See also Figure 1 and Figure 4 , an embodiment of the method of the present invention comprises the following steps:
[0026] Step 1: The actual handwritten signature dataset collected 500 A4 papers with handwritten signatures to test the accuracy of the improved capsule neural network in handwritten signature recognition. There are a total of 3500 signatures. First, the handwritten signature image is used as the image to be cut in the next step, and the handwritten signature information area to be recognized is cut and extracted to obtain a separate handwritten signature image, and then all images are unified to 224*224 size.
[0027] Step 2: Build an improved capsule neural network as a feature extractor
[0028] The structure of the improved capsule neural network: the input image of size 224*224 is first subjected to the classic convolution operation, and the data obtained by the convolution operation is used to construct the basic capsule. The basic capsule routes the data to the corresponding capsule in the next layer (handwriting capsule) through an adaptive clustering algorithm, and finally two fully connected layers are used to complete the image decoding and reconstruct the image.
[0029] Improved structure of capsule neural network Improvement point 1: Replace the dynamic routing algorithm in the traditional capsule neural network with an adaptive learning clustering algorithm to speed up the model operation and improve network performance.
[0030] See also Figure 2 ,The advantage of this method is that it removes the iterative process and reduces the ,amount of calculation., Since this method optimizes in a small range of predicted ,high-level feature vectors, it has a certain effect of adjusting the ,characterization morphological parameters, and can obtain a more accurate expression of the target object.
[0031] Improved structure of capsule neural network Improvement 2: In the reconstruction network of capsule neural network, Kronecker product is used to replace the fully connected layer, reduce the network model parameters, reduce the network dimension, and improve the network speed. For the schematic diagram of Kronecker product replacing the fully connected layer, see Figure 3 .
[0032] The number of parameters for a layer of Separable is n1m1+n2m2+n1n2, and the number of parameters for a layer of FC is n1m1×n2m2+n1n2. If the influence of the offset number is not considered for both, the parameter ratio can be expressed as:
[0033]
[0034] Because n1,n2,m1,m2>>1, we can infer Therefore, the Separable layer can greatly reduce the number of parameters compared to the FC layer.
[0035] Step 3: Build a least squares support vector machine as a classifier.
[0036] Step 4: Train the improved network model on the target data set to obtain a trained handwritten signature recognition model.
[0037] The convolution kernel of the ordinary convolution layer is set to 256, the convolution kernel size is set to 9×9 pixels, the step size is set to 1 pixel, the activation function is selected as the ReLU function, and the input image passes through the convolution layer to obtain an output tensor of 20×20×256; the primary capsule layer consists of 32 capsules, the convolution kernel size is set to 9×9 pixels, the step size is set to 2 pixels, and a total of 256 9×9 convolution kernels are included. The input is a tensor of 20×20×256 and the output is a tensor of 6×6×8×32; the handwriting capsule layer contains 10 digital capsules, which is a tensor of 10×16.
[0038] Step 5: Use the features extracted by the capsule neural network to train the least squares support vector machine.
[0039] Step 6: Prediction: Make predictions on new unknown images based on the trained handwritten signature recognition model.
[0040] The method provided by the present invention uses an improved initialization method for the weights of the improved capsule neural network model, and then trains and tests the network. After 400 iterations, the recognition time of the method in this paper and the traditional capsule neural network is as follows: Figure 5 As shown in the figure, the recognition accuracy comparison curve of the method of the present invention and the traditional capsule neural network is as follows Figure 6 As shown, the present invention improves a deep learning method for handwritten signature recognition.
[0041] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.
Claims
1. An improved capsule neural network handwritten signature recognition method, characterized in that: The following steps are involved: Step 1: Obtain and preprocess the handwritten signature dataset; Step 2, building a basic model of the improved capsule neural network, in which an adaptive learning clustering algorithm is used for feature extraction and a Kronecker product method is used for parameter calculation; The structure of the improved capsule neural network: the input image of size 224*224 is first subjected to the classic convolution operation, and the data obtained by the convolution operation is used to construct the basic capsule. The basic capsule routes the data to the corresponding capsule in the next layer of handwriting capsules through an adaptive clustering algorithm. Finally, two fully connected layers are used to complete the image decoding and reconstruction. Step 3: Build a least squares support vector machine as a classifier; Step 4, training: use the preprocessed handwritten signature dataset as the sample library to train the improved capsule neural network and set several epochs until the training accuracy is stable; Step 5: Input the features extracted by the capsule neural network into the least squares support vector machine to train the least squares support vector machine; Step 6: Prediction: Make predictions for new unknown images based on the trained handwritten signature recognition model.
2. The improved capsule neural network handwritten signature recognition method according to claim 1, characterized in that: In step 2, the collected image is first preprocessed, including grayscale processing, tilt processing, and denoising processing, and then the handwritten signature is cut and extracted as the information area to be identified to obtain an image with only the handwritten signature.
Citation Information
Patent Citations
Method for identifying handwritten numbers based on convolutional neural network and support vector machine
CN106599941A
Gait recognition method based on feedback weight convolutional neural network and capsule neural network
CN110110668A
Test question shooting method and device in click-to-read scene, electronic equipment and storage medium
CN111753715A
Method and apparatus of open set recognition and a computer readable storage medium
US20190147336A1
