A braille character recognition method based on an accelerometer of a smart wristband device

Abstract

This invention relates to a Braille character recognition method based on the accelerometer of a smart wristband device, belonging to the field of mobile computing application technology. This invention identifies Braille characters by analyzing the acceleration signals of the smart wristband device when a user touches Braille. It utilizes only the triaxial accelerometer built into the smart wristband device to capture the wrist movement patterns of the wearer when touching different Braille characters, thereby recognizing different Braille characters and helping visually impaired individuals learn and read Braille. This method effectively separates the wrist movement caused by the Braille dots and extracts three types of acceleration features: classical statistical features, character complexity features, and fractal geometry features. It robustly identifies Braille characters under different contact surface materials and different user conditions using a weighted Naive Bayes classification algorithm and the most discriminative features. This method is simple and direct to operate and does not affect wrist function.

Description

Technical Field

[0001] This invention relates to a method for intelligent recognition of Braille characters, specifically a method for recognizing Braille characters based on the accelerometer of a smart wristband device, belonging to the field of artificial intelligence and mobile computing application technology. Background Technology

[0002] For people with visual impairments, the lack of smooth communication and independence presents numerous difficulties and challenges in their lives and work. Braille, a tactile writing system designed specifically for the visually impaired, uses a combination of raised dot-shaped dots (also known as raised dots) in six positions across three rows and two columns to form individual characters. Braille improves the communication environment for the visually impaired and gives them greater independence in their lives. However, due to the high cost and time-consuming learning process, the proportion of visually impaired people who have the opportunity to learn Braille is declining, leading to a Braille literacy crisis. Therefore, a new solution is urgently needed to help visually impaired people read and learn Braille.

[0003] Currently, researchers are experimenting with various devices for automated Braille recognition. Based on the principles of perception, these studies fall into two categories: one based on computer vision technology (also known as optical Braille recognition), and the other based on tactile sensor technology.

[0004] Methods based on computer vision technology primarily utilize scanners and cameras to capture images of Braille documents. By analyzing the geometric structure of the Braille characters, these methods can convert the images into natural language characters. However, these methods rely on demanding image acquisition conditions, such as maintaining a specific height and angle between the camera and the Braille document, which is difficult for visually impaired individuals. Methods based on tactile sensor technology involve making gloves or attaching them to fingers using specialized materials (such as polyvinylidene fluoride film) to sense changes in pressure or electrical information when the user's fingers touch different Braille characters, ultimately achieving Braille character recognition. However, these materials are expensive and unsuitable for widespread deployment. Furthermore, these methods can affect other hand functions, causing inconvenience in daily life.

[0005] Therefore, existing methods have various defects and shortcomings, and new methods are urgently needed to solve the problem of Braille character recognition. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings and defects of existing technologies and to solve the technical problem of the lack of low-cost and easy-to-operate automatic Braille character recognition. It creatively proposes a Braille character recognition method based on the accelerometer of a smart wristband device. This method is simple, easy to operate, and does not affect hand function.

[0007] The innovation of this invention lies in the fact that different Braille characters have different raised dot arrangements. Therefore, when a user touches Braille characters while wearing a smart wristband device (such as a smartwatch), it triggers different hand movement patterns, which in turn affect the three-axis accelerometer of the smart wristband device. This invention identifies Braille characters by analyzing the acceleration signals of the smart wristband device when the user touches Braille. It utilizes the built-in three-axis accelerometer of the smart wristband device to capture the wrist movement patterns of the wearer when touching different Braille characters, thereby recognizing different Braille characters and helping visually impaired people learn and read Braille.

[0008] This invention is achieved through the following technical solution:

[0009] A method for recognizing Braille characters based on the accelerometer of a smart wristband device includes the following steps:

[0010] Step 1: Use the three-axis accelerometer in the smart wristband device to collect information on the motion state of the user's wrist when reading Braille, and remove random noise from the sensor and noise caused by other body movements.

[0011] Specifically, when a user reads Braille, an accelerometer worn on the wrist collects acceleration data along the X, Y, and Z axes, which are respectively recorded as a. x ,a y ,a z .

[0012] Step 1.1: Remove a x a y a z Random noise introduced by the accelerometer hardware can be removed, specifically using a low-pass filter or similar method.

[0013] Step 1.2: Process the triaxial signal a after step 1.1 x a y a z Denoising is performed (the purpose of which is to remove noise caused by other body movements) to obtain the motion signal caused by the Braille dots, called the dot signal, denoted as Da. x Da y Da z .

[0014] In this step, a denoising method based on wavelet packet decomposition and reconstruction is preferred.

[0015] Step 2: Segment the convex dot signal segments in the acceleration signal that correspond to the touch of Braille characters, and extract effective features from these segments. The purpose is to extract effective features from the convex dot signal segments of Braille characters to achieve Braille character recognition that is independent of the user and the contact surface material.

[0016] Specifically, step 2 may include the following steps:

[0017] Step 2.1: Calculate the linear acceleration of the convex signal obtained in Step 1.2, and apply a dynamic threshold to identify the start and end times of each Braille character. Based on the start and end times, in continuous Da... x ,Da y ,Da z In the signal, the convex dot signal segment corresponding to each character is segmented.

[0018] Step 2.2: Extract classical statistical features, character complexity features, and fractal geometry features from the convex dot signals corresponding to each character obtained in Step 2.1, and use them to construct a classification model.

[0019] Step 3: Using the features extracted from the convex signal segments, construct a weighted Naive Bayes classifier to recognize Braille characters.

[0020] Specifically, step 3 may include the following steps:

[0021] Step 3.1: Discretize the features obtained in Step 2.2, remove redundant features, select the most distinctive features, and satisfy the independence assumption of the classifier.

[0022] Step 3.2: Construct a weighted Naive Bayes classification model and perform high-precision Braille character recognition based on this model. That is, use a weighted Naive Bayes classifier to obtain the predicted probability of each convex signal segment belonging to 26 Braille characters, and select the predicted category with the highest probability as the Braille character to be recognized.

[0023] Beneficial effects

[0024] This method has the following advantages compared with the prior art:

[0025] 1. This invention achieves automatic and accurate recognition of Braille characters using only the accelerometer commonly found in smart wristband devices, resulting in low cost and simple operation. Furthermore, this invention captures the unique wrist movement patterns as the user reads different Braille characters, without interfering with the user's normal tactile experience.

[0026] 2. The method of this invention can effectively separate wrist movements caused by Braille bumps and extract three types of acceleration features, including classical statistical features, character complexity features, and fractal geometry features. Furthermore, this invention robustly identifies Braille characters under different contact surface materials and different user conditions using a weighted Naive Bayes classification algorithm and the most discriminative features. Attached Figure Description

[0027] Figure 1 It consists of 26 Braille characters arranged in a raised dot pattern;

[0028] Figure 2 This is a schematic diagram of the method of the present invention;

[0029] Figure 3 This is a schematic diagram of the wrist noise removal method of the present invention, wherein, Figure 3 a represents the convex signal contaminated by body movement noise; Figure 3 b represents the clean convex signal obtained after noise reduction;

[0030] Figure 4 The performance of the present invention in recognizing Braille characters;

[0031] Figure 5 This invention demonstrates the performance of Braille character recognition in 8 volunteers in an embodiment of the invention.

[0032] Figure 6 This invention demonstrates the performance of Braille character recognition under different contact surface material conditions in embodiments of the present invention.

[0033] Figure 7 This invention provides a long-term performance test for Braille character recognition. Detailed Implementation

[0034] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0035] In this specific embodiment, the standard Braille characters of the 26 English letters A, Z, and E are used as the specific detection target. Each Braille character is formed within a Braille unit, which contains 6 dots arranged in 3 rows and 2 columns. Figure 1 The Braille representation of the 26 English letters is displayed, where solid indicates raised dots and hollow indicates no raised dots.

[0036] like Figure 2 As shown, a Braille character recognition method based on the accelerometer of a smart wristband device includes the following steps:

[0037] Step 1: Use the three-axis accelerometer in the smart wristband device to collect information on the motion state of the user's wrist when reading Braille, and remove random noise from the sensor and noise caused by other body movements.

[0038] Specifically, when a user reads Braille, an accelerometer worn on the wrist collects acceleration data along the X, Y, and Z axes, which is recorded as a. x ,a y ,a z .

[0039] Step 1.1: Use a low-pass filter to remove a x ,a y ,a zRandom noise introduced by the accelerometer hardware.

[0040] The frequency of human finger movements is generally below 12Hz, and the hardware itself introduces high-frequency random noise. Therefore, in this embodiment, a dual-pass fourth-order Butterworth filter is used to filter a. x ,a y ,a z Low-pass filtering (specifically, 12Hz) is applied to remove random noise.

[0041] Step 1.2: Using a denoising method based on wavelet packet decomposition and reconstruction, the triaxial signal a processed in Step 1.1 is denoised. x ,a y ,a z Denoising is performed separately to remove noise caused by other body movements and obtain the motion signal caused by the Braille dots, called the dot signal, denoted as Da. x Da y Da z

[0042] Figure 3 a shows the data captured from the wrist while reading Braille characters. x a y ,a z Signal example. Specifically, between seconds 6 and 8, the wearer's hand suddenly trembles, a y and a z Significant motion noise was generated in the signal. During other times, the wearer maintained normal Braille reading with stable hand movements. This demonstrates that in the presence of motion noise, the acceleration signal was severely distorted, and the noise dominated the signal amplitude. This noise interference is random and unpredictable, affecting the accurate recognition of subsequent Braille characters, and therefore needs to be eliminated.

[0043] Specifically, first, select the wavelet packet basis and the number of decomposition levels N to pair a x a y a z The three-axis signals are decomposed into wavelet packets, and the signal of each axis is obtained as 2. N Sub-band. In this embodiment, the DB6 wavelet envelope basis is selected; however, the setting of other wavelet envelope bases belonging to the DB series is also within the scope of this invention.

[0044] When a signal is decomposed using wavelet packets, the effective number of decomposition levels depends on the motion signal and the sampling frequency. Since the frequency of wrist motion signals is typically less than 12Hz, based on Shannon sampling, the sensor's sampling rate is set to 200Hz, indicating its Nyquist frequency is 100Hz. According to the wavelet packet decomposition principle, the signal can be decomposed into three levels, i.e., N=3, for a total of 2... 3= 8 wavelet packet subbands. Each subband width is 12.5(100 / 8)Hz, approximately 12Hz, which will not cause signal distortion.

[0045] Then, calculate a according to Equation 1 in turn. x ,a y ,a z Cross-correlation values ​​of the three-axis signal and its eight sub-bands:

[0046]

[0047] Where, ρ e Indicate M and M e The cross-correlation value, and These represent the original signal M (representing a in sequence). x ,a y ,a z ) and the e-th subband M e The average value. Experiments show that the closer the cross-correlation value is to 1, the closer the signals are to each other. e A value greater than 0.5 indicates that the two signals are indeed correlated.

[0048] Then, based on the calculated cross-correlation values, the subbands are grouped and processed.

[0049] Specifically, ρ e Subbands with a correlation coefficient ≥ 0.5 are denoted as Group A, and these subbands are directly considered as motion noise components because the amplitude of the noise is much larger than that of the pure convex signal. After decomposition, the noise components will fall into the subbands with high correlation values. For subbands with a correlation coefficient ≥ 0.1, the following subbands are considered: e Subbands with a value <0.5 are designated as Group B and are considered as half-noise components. The remaining subbands are designated as Group C and are considered as noise-free components.

[0050] Then, for the B group subband, the threshold λ is calculated using the Stein adaptive thresholding rule, and wavelet soft thresholding denoising is applied for noise reduction. The soft thresholding function is defined as follows:

[0051]

[0052] Where, ψ k and These are the original wavelet packet coefficients and the threshold wavelet packet coefficients, respectively; k is the wavelet packet node; sgn(ψ) k ) is the sign function; λ is the threshold, calculated according to Equation 3:

[0053]

[0054] Where L is the length of the signal, med(|M e |) is the median of the wavelet packet coefficients for each detail.

[0055] Finally, wavelet packet reconstruction is performed on the subbands in group C and the thresholded subbands in group B to obtain a clean, denoised motion signal caused by the convex points, called the convex point signal, denoted as Da. x ,Da y ,Da z .

[0056] The final denoising result is as follows Figure 3 As shown in (b).

[0057] Step 2: Segment the convex dot signal segments in the acceleration signal that correspond to the touch of Braille characters, and extract effective features from these segments. The purpose is to extract effective features from the convex dot signal segments of Braille characters to achieve Braille character recognition that is independent of the user and the contact surface material.

[0058] Specifically, step 2 includes the following steps:

[0059] Step 2.1: Calculate the linear acceleration of the convex signal obtained in Step 1.2, and apply a dynamic threshold to identify the start and end times of each Braille character. Based on the start and end times in continuous Da... x ,Da y ,Da z The signal is segmented into the convex dot signal segment corresponding to each character.

[0060] To eliminate the influence of device coordinate system rotation based on wearing direction, this invention uses the amplitude of linear acceleration to guide the segmentation of the three-axis convex point signal. The linear acceleration MLA is shown in Equation 4:

[0061]

[0062] Among them, Da x Da y and Da z These are the readings on the X, Y, and Z axes of the accelerometer after noise reduction; g is the gravitational acceleration.

[0063] Then, the start and end times of each Braille character touch are detected using a short-time energy zero-crossing ratio, thereby segmenting the convex dot signal. Because the energy value during a touch is much higher than the zero-crossing rate value compared to a period without any movement, this invention divides the energy value by the zero-crossing rate value, amplifying the difference between the Braille character touch and a period without any movement, making it easier to detect the start and end points.

[0064] First, the MLA signal is framed using a 200ms sliding window with 100ms overlap.

[0065] Then, the zero ratio EZR(m) is:

[0066]

[0067] Where STE(m) and ZCR(m) are the short-time energy and zero-crossing rate of the m-th frame, respectively. The threshold is set to twice the average of the energy-to-zero ratio during the 100ms static period before the touch motion. Continuous time windows exceeding the threshold are considered as Braille character bump signal segments.

[0068] Step 2.2: Extract classical statistical features, character complexity features, and fractal geometry features from the convex dot signals corresponding to each character obtained in Step 2.1 to construct the subsequent classification model.

[0069] The classic statistical characteristics are as follows:

[0070] Descriptive statistics for each convex signal segment are extracted to effectively study the entire dataset and reflect the statistical characteristics of unique wrist movement patterns associated with Braille characters. Central tendency and variance are two concepts describing the data. The mean is extracted from the central tendency as a measure of centrality, while variance and standard deviation are extracted from the variance to characterize the stability of the motion signal; these are considered fundamental features. Furthermore, commonly used time-domain and frequency-domain features such as the mean, mean deviation, standard deviation, minimum, and maximum values ​​of linear acceleration can be used to further characterize the unique wrist movement patterns.

[0071] The character complexity characteristics are as follows:

[0072] To explore the potential correlation between Braille character complexity and wrist movement patterns, this invention constructs a quantitative representation of complexity, linking objective measurement with perceptual complexity. Specifically, the complexity γ of a Braille character is calculated using the number of bumps (Dot) and the subsymmetry number (Sym) in the character unit:

[0073] γ=σ·Dot+(1-σ)·Sym (6)

[0074] Where σ represents the weighting coefficient, and the value of σ ranges from [0,1]. In this embodiment, it is set to 0.94. The Sym function is derived from the following equation:

[0075]

[0076] Where X and Y are 3 and 2 respectively, representing the number of rows and columns of Braille characters. N (p,q) (x, y) is the number of symmetries observed at position (p, q) within a rectangle of size x × y, where x + y > 2; symmetry includes vertical, horizontal, and central symmetry across all subgrid sizes, as well as diagonal symmetry when x = y. Sym counts the number of symmetries for each rectangular subgrid in the Braille character cell matrix and calculates their weighted sum, with the weight being the area of ​​the subgrid.

[0077] Finally, we explored features related to the complexity of Braille characters. By calculating and averaging candidate features for each character in 200 experiments, and calculating their cross-correlation values ​​with complexity γ according to Equation 1, we found that the cross-correlation values ​​of zero-crossing rate, slope sign change, and wavelength were all greater than 0.5, indicating that they are correlated with the complexity of Braille characters. Therefore, they were selected as the basic character complexity features.

[0078] The characteristics of fractal geometry are as follows:

[0079] Wrist motion signals exhibit fractal characteristics. Furthermore, inspired by optical Braille recognition, fractal analysis was performed on the signals, extracting the Higuchi and Katz fractal dimensions as fractal geometric features. These reflect the effectiveness of the space occupied by complex shapes and are measures of the irregularity of complex shapes.

[0080] Step 3: Using the features extracted from the convex signal segment, construct a weighted Naive Bayes classifier to recognize Braille characters.

[0081] Specifically, it includes the following steps:

[0082] Step 3.1: Discretize the features obtained in Step 2.2, remove redundant features, select the most distinctive features, and satisfy the independence assumption of the classifier.

[0083] The extracted three types of features are subjected to correlation removal, and the most distinctive features are selected. While these features effectively characterize the convex signal, they have potential correlations. When all features are input into the classifier, these correlations can negatively impact its performance. Therefore, feature selection is necessary to find the optimal subset, thus satisfying the classifier's independence assumption.

[0084] First, the ChiMerge algorithm is used to discretize continuous feature values. Then, to satisfy the independence assumption, the most discriminative features are automatically selected from the extracted features. Specifically, symmetric uncertainty is used to measure the correlation between features and between features and Braille character classes, and redundant features are eliminated according to a fast correlation filtering algorithm.

[0085] Symmetric uncertainty is an effective method for measuring the correlation between two variables, with values ​​ranging from [0,1]. A larger value indicates a higher correlation. The characteristic F is calculated. i The symmetric uncertainty value SU of Braille character class C Fi,C The details are as follows:

[0086]

[0087] Among them, H(F)i H(C) and H(C) are characteristic F i And the information entropy of category C; IG(F i |C) is the feature F under category C. i The increased information content, or information gain, is calculated as follows:

[0088] IG(F i |C)=H(F) i )-H(F i |C)=H(F) i )-H(C)+H(F i ,C) (9)

[0089] Among them, H(F) i |C) represents feature F i The conditional entropy between class C and category C, H(F) i C) represents feature F i The joint entropy between class C and class C.

[0090] Similarly, feature F i and feature F j Symmetric uncertainty values ​​between The calculation is as follows:

[0091]

[0092] Where i≠j, H(F) j ) represents feature F j Information entropy.

[0093] Sort the features and categories C in descending order of their symmetric uncertainty. Select the top-ranked feature and eliminate its redundant features using a fast correlation filtering algorithm. Then, select the next ranked feature for filtering until all redundant features are eliminated.

[0094] Step 3.2: Construct a weighted Naive Bayes classification model and perform high-precision Braille character recognition based on this model. That is, use a weighted Naive Bayes classifier to obtain the predicted probability of each convex signal segment belonging to 26 Braille characters, and select the predicted category with the highest probability as the Braille character to be recognized.

[0095] Specifically, since different features have different classification abilities, it is necessary to measure each feature based on the degree of association between the feature and the category, as well as the correlation between features, in order to distinguish the importance of each feature to the classification.

[0096] Weight of each feature Calculated as:

[0097]

[0098] Where κ is the number of features after removing redundant features in step 3.1.

[0099] Then, the weights are normalized:

[0100]

[0101] in, This represents the normalized weights.

[0102] Then the weighted Naive Bayes model P(C j ||s) is:

[0103]

[0104] Among them, s i It is the feature set of the convex signal segment during prediction; j is the character category, and

[0105] Finally, calculate P(C) j ∣∣s), choose class C with the highest probability. j , as a class for Braille character recognition.

[0106] Example

[0107] To verify the performance of the invention, a wristband prototype was developed, comprising an integrated inertial motion sensor containing a three-axis accelerometer, equivalent to a smartwatch or smart bracelet. During reading, accelerometer readings were recorded at a sampling rate of 200Hz, and a camera was placed in front of the volunteers to provide a reference for real-world movements. This experiment recruited eight volunteers (four men and four women, aged 20 to 35 years). All volunteers were healthy, right-handed, and had no prior experience with Braille reading. After brief training, volunteers were asked to wear the prototype and bezel-covered. The experiment used six Braille documents on standard Braille paper, and in subsequent experiments, Braille characters on materials such as cardboard and stainless steel were used, collecting over 500 fragments for each of the 26 Braille characters.

[0108] Recall, precision, and F1 score are used to evaluate system performance. Recall is defined as the percentage of segments correctly identified as character type A out of all segments predicted to be of character type A; precision is defined as the percentage of segments correctly identified as type A out of all segments predicted to be of character type A; and F1 score is defined as:

[0109] First, the overall performance of this invention in recognizing 26 Braille characters was tested. 80% of the data was used for training, and the remaining 20% ​​was used for testing, with five rounds of cross-validation. For each character, the average value among different users was calculated. Figure 4 The experimental results are presented, showing an average recall of 91.65%, an average precision of 91.52%, and an average F1 score of 91.58%. Furthermore, relatively low recall and precision can be observed for some characters, such as "H", "S", "T", and "Z". This may be because the convex signal segments of "H", "S", and "T" have similar structural complexity values ​​and speedup zero-crossing, causing most features to fail. "Z" occurs less frequently among characters, resulting in a smaller training data size for it compared to other characters, thus reducing accuracy.

[0110] To assess the user independence of this invention, it was tested using data from one volunteer and trained using data from the remaining seven volunteers. Figure 5 The average recall, precision, and F1 score for all characters for each volunteer are displayed. It can be observed that the average recall, precision, and F1 score are 91.95%, 90.92%, and 91.43%, respectively. Meanwhile, seven volunteers achieved precision, recall, and F1 scores all above 90%. These results demonstrate that the present invention can work effectively across different users.

[0111] Besides books, Braille also appears on packaging boxes, elevator buttons, and other items. Therefore, in order to understand whether the invention could be extended to new materials without retraining, experiments were conducted on surfaces of three typical materials (cardboard, kraft paper, and stainless steel). Figure 6 The average recall, precision, and F1 score are shown. Kraft paper had the highest recall (94.91%), precision (93.86%), and F1 score (94.38%), while cardboard had the lowest recall (83.07%), precision (83.18%), and F1 score (83.12%). The results indicate that the present invention exhibits robust recognition performance across different surface materials.

[0112] Next, the long-term performance of the method of the present invention is tested. Long-term performance is a key aspect of cognitive performance because most people's behavioral patterns change slightly over time. Figure 7 The Braille character recognition performance of all volunteers over 30 days was described. After training, test data were collected on the same day, 1 day later, 2 days later, 5 days later, 10 days later, 20 days later, and 30 days later. When tested using data collected on day 30, the recall, precision, and F1 score of this invention were all above 88%, a result acceptable in a real-world environment.

[0113] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is used to explain the present invention. It is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for recognizing Braille characters based on the accelerometer of a smart wristband device, characterized in that, Using the 26 standard Braille characters of the English alphabet A, Z, as the specific detection target, the following steps are included: Step 1: Use the three-axis accelerometer in the smart wristband device to collect the motion status information of the user's wrist when reading Braille, and remove random noise from the sensor and noise caused by other body movements; When a user reads Braille, an accelerometer worn on the wrist collects acceleration data along the X, Y, and Z axes, which is recorded as a. x ,a y ,a z ; First, use a low-pass filter to remove a. x ,a y ,a z Random noise introduced by the accelerometer hardware; Then, using a denoising method based on wavelet packet decomposition and reconstruction, the three-axis signal a is denoised. x ,a y ,a z Denoising is performed separately to obtain the motion signal caused by the Braille dots, called the dot signal, denoted as Da. x ,Da y ,Da z ; Step 2: Segment the convex dot signal segments in the acceleration signal that belong to the touch of Braille characters, and extract effective features from the convex dot signal segments; Step 3: Using the features extracted from the convex signal segment, construct a weighted Naive Bayes classifier to recognize Braille characters; Step 3 includes the following steps: Step 3.1: Discretize the features obtained in Step 2.2, remove redundant features, select the most distinctive features, and satisfy the independence assumption of the classifier; First, the ChiMerge algorithm is used to discretize continuous feature values; then, the most discriminative features are automatically selected from the extracted features, symmetric uncertainty is used to measure the correlation between features and between features and Braille character classes, and redundant features are eliminated according to the fast correlation filtering algorithm. Symmetric uncertainty is an effective method for measuring the correlation between two variables, with a value range of [0,1]. A larger value indicates a higher correlation. The characteristic F is calculated. i Symmetric uncertainty value with respect to Braille character class C Specifically as follows: Among them, H(F) i H(C) and H(C) are characteristic F i And the information entropy of category C; IG(F i |C) is feature F under category C. i The increased information content, or information gain, is calculated as follows: IG(F i |C)=H(F i )-H(F i |C)=H(F i )-H(C)+H(F i ,C) (9) Among them, H(F) i |C) represents feature F i The conditional entropy between class C and category C, H(F) i C) represents feature F i Joint entropy between and category C; Feature F i and feature F j Symmetric uncertainty values ​​between The calculation is as follows: Where i≠j, H(F) j ) represents feature F j Information entropy; Sort the features and categories C in descending order of their symmetric uncertainty. Select the top-ranked feature and eliminate its redundant features using a fast correlation filtering algorithm. Then, select the next ranked feature for filtering until all redundant features are eliminated. Step 3.2: Construct a weighted Naive Bayes classification model and perform high-precision Braille character recognition based on the model. Use a weighted Naive Bayes classifier to obtain the predicted probability of each convex signal segment belonging to 26 Braille characters, and select the predicted category with the highest probability as the Braille character to be recognized. Each feature is measured by the degree of association between features and categories, as well as the correlation between features. The weight of each feature is determined by its relationship to the category. Calculated as: Where κ is the number of features after removing redundant features in step 3.1; Then, the weights are normalized: in, This represents the normalized weights; Then the weighted Naive Bayes model P(C j |s) is: Among them, s i It is the feature set of the convex signal segment during prediction; j is the character category, and Finally, calculate P(C) j |s), choose class C with the highest probability. j , as a class for Braille character recognition.

2. The Braille character recognition method based on the accelerometer of a smart wristband device as described in claim 1, characterized in that, In step 1, firstly, the wavelet envelope basis and the number of decomposition levels N are selected for a. x ,a y ,a z The three-axis signals are decomposed into wavelet packets, and the signal of each axis is obtained as 2. N Individual belt; Then, calculate a according to Equation 1 in turn. x ,a y ,a z Cross-correlation values ​​of the three-axis signal and its eight sub-bands: Where, ρ e Indicate M and M e The cross-correlation value, and Let M represent the original signal M and the e-th subband M, respectively. e On average, when ρ e A value greater than 0.5 indicates that the two signals are indeed correlated; Next, based on the calculated cross-correlation values, the subbands are grouped and processed, and ρ is... e Subbands with a correlation coefficient ≥ 0.5 are denoted as Group A, and these subbands are directly considered as motion noise components because the amplitude of the noise is much larger than that of the pure convex signal. After decomposition, the noise components will fall into the subbands with high correlation values; 0.1 < ρ e Subbands with a value <0.5 are designated as Group B and are considered as half-noise components; the remaining subbands are designated as Group C and are considered as noise-free components. Then, for the B group subband, the threshold λ is calculated using the Stein adaptive thresholding rule, and wavelet soft thresholding denoising is applied for denoising; whereby the soft thresholding function is defined as: Where, ψ k and These are the original wavelet packet coefficients and the threshold wavelet packet coefficients, respectively; k is the wavelet packet node; sgn(ψ) k ) is the sign function; λ is the threshold, calculated according to Equation 3: Where L is the length of the signal, med(|M e |) is the median of the wavelet packet coefficients for each detail; Finally, wavelet packet reconstruction is performed on the subbands in group C and the thresholded subbands in group B to obtain a clean, denoised motion signal caused by the convex points, called the convex point signal, denoted as Da. x ,Da y ,Da z .

3. The Braille character recognition method based on the accelerometer of a smart wristband device as described in claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Calculate the linear acceleration of the convex signal obtained in Step 1, and apply a dynamic threshold to identify the start and end times of each Braille character; based on the start and end times, in continuous Da... x ,Da y ,Da z The signal is segmented to extract the dotted signal segment corresponding to each character; Step 2.2: Extract classical statistical features, character complexity features, and fractal geometry features from the convex dot signals corresponding to each character obtained in Step 2.1 to construct the subsequent classification model; Among them, the classical statistical features include: the concentration and the difference of the convex signal segment. The mean is extracted from the concentration as a measure of central tendency, and the variance and standard deviation are extracted from the difference to characterize the stability of the motion signal. These are regarded as the basic features. Character complexity features include: zero-crossing rate, slope sign change, and cross-correlation value of wavelength; Fractal geometry features include the Higuchi fractal dimension and the Katz fractal dimension.

4. The Braille character recognition method based on the accelerometer of a smart wristband device as described in claim 3, characterized in that, In step 2.1, the amplitude of linear acceleration is used to guide the segmentation of the triaxial convex signal. The linear acceleration MLA is shown in Equation 4: Among them, Da x Da y and Da z These are the readings of the accelerometer on the X, Y, and Z axes after noise reduction; g is the acceleration due to gravity. Then, the start and end times of each Braille character touch are detected using the short-time energy zero ratio, thereby segmenting the convex dot signal; the energy value is divided by the zero-crossing rate value to amplify the difference between the Braille character touch and the period without any movement. First, the MLA signal is framed using a 200ms sliding window with 100ms overlap; The zero-ratio EZR(m) is: Wherein, STE(m) and ZCR(m) are the short-time energy and zero-crossing rate of the m-th frame, respectively; the threshold is set to twice the average value of the zero-crossing rate of the 100ms static segment before the touch motion occurs; continuous time windows above the threshold are considered as Braille character bump signal segments.

Images