Chinese annotation gesture recognition method and system based on improved YOLOv5

By improving the FPN structure and fast normalized fusion strategy of the YOLOv5 algorithm, the problem of low accuracy and robustness in intelligent driving gesture recognition is solved, and the recognition of different body shapes and hand positions is realized, and Chinese labeling is supported to improve the The breadth of application in the Chinese market.

CN114332923BActive Publication Date: 2024-11-08YANGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111551115.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-17
Publication Date
2024-11-08
Estimated Expiration
2041-12-17

AI Technical Summary

Technical Problem

The existing gesture recognition method based on YOLOv5 is low in accuracy and robustness in intelligent driving, and cannot adapt to the differences in body shapes and hand distances and angles of different drivers. The label is displayed in English, which cannot meet the usage needs of the Chinese market. .

Method used

Improve the FPN structure of the YOLOv5 algorithm, introduce a fast normalized fusion strategy, increase data set diversity and image augmentation, use Chinese annotation tools to simplify the FPN structure and perform bidirectional cross-scale connections, and improve feature fusion efficiency and recognition accuracy.

Benefits of technology

It improves the accuracy and robustness of gesture recognition, adapts to the recognition of different drivers' body shapes and hand positions, supports Chinese annotation, and expands the application scope.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114332923B_ABST
    Figure CN114332923B_ABST
Patent Text Reader

Abstract

The present invention discloses a Chinese-annotated gesture recognition method and system based on improved YOLOv5, which can be used for intelligent driving gesture control. The present invention first performs image augmentation on the selected 0 to 5 six gesture training data sets; then performs preprocessing operations on the data set, using piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection operations to obtain the target image; then adds specific command meanings to the preprocessed target data set, improves the FPN structure of the YOLOv5 algorithm, introduces a fast normalization fusion strategy, and uses the improved algorithm to train the obtained target data set; after the training is completed, target detection is performed on the image to be detected, and the Chinese label and recognition accuracy are displayed above the image. Compared with the prior art, the present invention adopts an efficient feature fusion method, which can improve the recognition accuracy, reduce the occurrence of overfitting, and realize Chinese character display by marking with Chinese labels, thereby increasing the breadth of application.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, target detection and recognition, and in particular to a Chinese annotation gesture recognition method and system based on improved YOLOv5. Background Art

[0002] Gestures are basic human movement characteristics. Different gestures contain different characteristics and can represent special meanings. At the same time, with the development of society, traditional hardware input devices can no longer meet people's needs for human-computer interaction. Gesture recognition has become a key technology in the field of human-computer interaction. In particular, gesture recognition is more widely used in intelligent driving of Internet of Vehicles. Drivers can realize related intelligent functions through customized gestures, thereby reducing distraction, focusing more on road conditions and improving driving safety.

[0003] There are currently three main methods for intelligent driving gesture recognition on the market. The first is to use special wearable devices, but the operation is complicated, customized equipment is relatively expensive, and there are hardware contact problems, which makes it difficult to be widely used; the second is to use traditional gesture segmentation methods to train the effective part after segmentation in the classifier. Due to the complex gesture environment, it is difficult to accurately extract the effective area under the same method. Its robustness is poor and the accuracy is low, which is difficult to meet practical applications; the third is to use deep learning methods to train a large number of test sets to determine the gesture category. This method has high requirements for the data set and requires representative and universal images. The training time is long, but it can improve the recognition accuracy.

[0004] Compared with the above methods, small target detection represented by the YOLO algorithm based on deep learning convolutional neural network has obvious advantages. YOLO solves object detection as a regression problem, with high detection efficiency, low background false detection rate and strong versatility. Among them, YOLOv5 has good performance in model flexibility, rapid deployment and convergence. YOLOv5 consists of four parts: input, Backbone, Neck and Predication. The newly added Focus module of Backbone uses slicing operation to layer the target image, expands the input channel to 4 times the original, and reduces the amount of calculation after one convolution. Neck adopts a structure combining FPN and PAN, combines FPN with feature pyramid, integrates semantic features and position features, and uses GIOU_Loss as the loss function of Bounding box at the output to reduce deviation.

[0005] The gesture recognition method based on YOLOv5 currently used in intelligent driving has a relatively complex structure and cannot take into account practical problems related to the distance and angle of the hand from the image collector and the fatness or thinness of the hand due to the different body shapes of drivers. In addition, it is limited by the low-pixel and small-size images collected by the image collector, which brings great difficulty to the training process and results in low accuracy and robustness. In the result display diagram of algorithm recognition, its annotation labels and target result labels are displayed in English and cannot support Chinese characters, which is inconsistent with the usage habits of Chinese people. Therefore, it is difficult to be promoted in the Chinese market, resulting in a limited range of users. Summary of the invention

[0006] Purpose of the invention: In view of the shortcomings in the prior art, the purpose of the present invention is to provide a Chinese annotation gesture recognition method and system based on improved YOLOv5, and to achieve a more efficient feature fusion method by improving FPN. Since different input features have different resolutions and their contributions to the output features are unequal, a fast normalized fusion strategy is introduced to improve the recognition accuracy, reduce the occurrence of overfitting, and reduce the misrecognition of the number of fingers in the gesture to be recognized due to the influence of the distance and angle between the driver's hand and the image collector and the fatness and thinness of the hand.

[0007] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution: a Chinese annotation gesture recognition method based on improved YOLOv5, comprising the following steps:

[0008] Step 1: Select a training data set, which includes 0 to 5 gestures of different people. It is necessary to shoot gestures of different people, including men and women, fat and thin, angle changes, etc., and add interference gestures;

[0009] Step 2: Augment the obtained gesture dataset by performing image augmentation, including operations such as rotation affine transformation, brightness transformation, adding Gaussian noise, and any combination of operations in any order;

[0010] Step 3: Perform preprocessing operations on the obtained data set, using pre-operations such as piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection to obtain the target image, which constitutes the target data set;

[0011] Step 4: Use the modified LabelImg annotation tool to annotate the preprocessed target dataset, add specific command meanings to each gesture from 0 to 5, and obtain the target dataset with the suffix .xml;

[0012] Step 5: Improve the FPN structure of the YOLOv5 algorithm, introduce a fast normalization fusion strategy, and use the improved algorithm to train the obtained target data set;

[0013] Step 6: After training is completed, target detection is performed on the image to be detected, and the Chinese label and recognition accuracy are displayed above the image.

[0014] The present invention can complete the tasks of solving practical problems related to the distance of the hand from the image collector, different angles, fat or thin hand shapes, etc. under different backgrounds in the car according to the different body shapes of drivers, and give static gestures 0 to 5 specific command meanings, add Chinese labels to images, and display recognition accuracy labels, so as to achieve the purpose of improving the accuracy and robustness of gesture recognition.

[0015] Preferably, only three feature layers of 76*76, 38*38, and 19*19 are retained on the original FPN structure for feature fusion, and the two intermediate nodes of the 19*19 feature layer and the 76*76 feature layer are omitted. The 19*19 feature layer of the improved FPN structure is upsampled and fused with the feature layer of the intermediate dimension of 38*38. The fused output result is then connected to the 76*76 feature layer, and then a button-up bottom-up channel is added. At the same time, a jump connection of the residual edge connecting the input feature map and the output feature map is added to the feature layer with an output size of 38*38. This greatly simplifies the FPN structure and realizes efficient bidirectional cross-scale connection and multi-scale fusion, thereby improving the efficiency of the algorithm.

[0016] Preferably, a fast normalized fusion method is introduced in terms of weights. In some scenarios with large data differentiation, the data is uniformly mapped to [0,1] to reduce numerical instability. Compared with the traditional softmax method, the operation of defining weights as probabilities is changed. This method introduces nonlinear factors into neurons, improves the speed and convergence speed of the gradient descent method to find the optimal solution, and is more widely used and more efficient. It is applied to the weight of each input feature to achieve a normalization effect, and the speed is greatly improved. The weight formula is as follows:

[0017]

[0018] Output is the output image result after fast normalization, w i is a learnable weight, smooth is a small amount that ensures numerical stability, i and j are the feature map indices input to the feature fusion node, and I i is the input feature map matrix. In order to ensure that w i≥0, where the ReLU activation function can overcome the gradient vanishing problem and speed up the training. The output f(x) function formula of the input vector x from the previous layer of the neural network entering the neuron after using the ReLU function is:

[0019] f(x)=max(0,x)

[0020] Preferably, the Windows version of LabelImg image annotation tool is modified to change the character type, call the Windows internal Chinese font, and use variable-length character encoding to realize Chinese character display, so that the open file path can contain Chinese directories, and the functions of annotating labels in Chinese and displaying Chinese labels can also be realized.

[0021] Preferably, in step 2, the rotated image is filled with black to reduce the amount of training calculations and improve detection efficiency.

[0022] Based on the same inventive concept, the present invention provides a Chinese annotation gesture recognition system based on improved YOLOv5, including the following modules:

[0023] An acquisition module is used to obtain a training data set, which includes six gestures (0 to 5) of different people;

[0024] The image augmentation module is used to augment the original data set, including rotation affine transformation, brightness transformation, adding Gaussian noise operation, and any combination of operations in any order;

[0025] The image preprocessing module is used to perform preprocessing operations on the obtained data set, using piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection operations to obtain the target image, which constitutes the target data set;

[0026] The data set annotation module is used to annotate the preprocessed target data set using the modified LabelImg annotation tool, adding specific command meanings to each gesture from 0 to 5, and obtaining the target data set with the suffix .xml;

[0027] The model training module is used to improve the FPN structure of the YOLOv5 algorithm and introduce a fast normalization fusion strategy, and use the improved algorithm to train the obtained target data set;

[0028] As well as the gesture recognition module, it is used to perform target detection on the image to be detected after training is completed, and display the Chinese label and recognition accuracy above the image.

[0029] Based on the same inventive concept, the present invention provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the Chinese annotation gesture recognition method based on the improved YOLOv5 is implemented.

[0030] Beneficial effects: Compared with the prior art, the advantages of the present invention are:

[0031] 1. Simplify the YOLOv5 network structure. Only three feature layers of 76*76, 38*38, and 19*19 are retained on the original FPN structure for feature fusion, and the two intermediate nodes of the 19*19 feature layer and the 76*76 feature layer are omitted. The 19*19 feature layer of the improved FPN structure is upsampled and fused with the feature layer of the 38*38 intermediate dimension. The fused output result is connected with the 76*76 feature layer, and then a button-up bottom-up channel is added. At the same time, a jump connection of the residual edge connecting the input feature map and the output feature map is added to the feature layer with an output size of 38*38. This greatly simplifies the FPN structure and realizes efficient bidirectional cross-scale connection and multi-scale fusion, thereby improving the efficiency of the algorithm.

[0032] 2. Use the fast normalized fusion method to uniformly map the data to [0,1] in some scenarios with large data differentiation, reducing numerical instability. Compared with the traditional softmax method, this method changes the operation of defining weights as probabilities. This method introduces nonlinear factors into neurons, improves the speed and convergence of the gradient descent method to find the optimal solution, and is more widely used and more efficient.

[0033] 3. By augmenting the data set and the number of samples, we can achieve the actual effect under different interferences, improve the detection accuracy, and reduce the overfitting phenomenon. At the same time, we fill the rotated images with black to reduce the amount of training calculations and improve the detection efficiency.

[0034] 4. Preprocess the augmented data set, where the piecewise linear transformation increases the grayscale area of ​​interest in the image, the histogram equalization method averages the grayscale histogram to enhance the image clarity, the smoothing denoising method eliminates noise and smoothes edge information, and the edge detection uses the adaptive threshold method to achieve the contour recognition of the target gesture, reduce the amount of calculation, and improve the detection efficiency;

[0035] 5. Use the improved Labelmg annotation tool with variable-length character encoding to mark Chinese labels to realize Chinese character display, increase the breadth of application, give practical meaning to actions, be more in line with Chinese habits, and open up the market in China. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 FIG. 4 is a flowchart of gesture recognition according to an embodiment of the present invention.

[0037] Figure 2 Schematic diagram of the improved YOLOv5 algorithm in an embodiment of the present invention.

[0038] Figure 3 This is a diagram of the improved FPN structure in an embodiment of the present invention.

[0039] Figure 4 It is the function image of the ReLU activation function in the embodiment of the present invention.

[0040] Figure 5 This is an example diagram of a data set in an embodiment of the present invention.

[0041] Figure 6 This is an augmented example diagram in the embodiment of the present invention.

[0042] Figure 7 This is a result diagram of an image after preprocessing in an embodiment of the present invention.

[0043] Figure 8 This is a result diagram of the data set in the embodiment of the present invention marked with Chinese.

[0044] Fig. 9 4 is a diagram showing the detection result of the gesture to be detected in an embodiment of the present invention.

[0045] Fig.10 This is a visualization analysis diagram of training results in an embodiment of the present invention (x-axis: number of training rounds). DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0047] The present invention is based on the improved deep learning YOLOv5 algorithm. Figure 1 : is a flowchart of a gesture detection and recognition method according to an embodiment of the present invention, which comprises the following steps:

[0048] Step 1: Training dataset image screening

[0049] Select six international hand gestures from 0 to 5. Different photographers need to be selected to shoot. The data set needs to include multiple situations such as gender, fatness, thinness, angle changes, etc., and add interference gestures. Figure 5 An example graph of the dataset is shown.

[0050] Step 2: Dataset image augmentation

[0051] Through operations such as rotation affine transformation, brightness change, adding Gaussian noise, and combinations of any order, Figure 6 An augmented example graph is shown.

[0052] (1) Use the rotational affine transformation method to increase the data set

[0053] Assume that the coordinates of a given point in the original image are (x0, y0), and the corresponding point coordinates after the rotational affine transformation are (x, y), the image with a height of H and a width of W, after rotation by angle θ, the height and width of the image contained correspond to H' and W' respectively:

[0054] ① Convert the input original image coordinates into the Cartesian coordinate system, as shown in the formula (1):

[0055]

[0056] ② Perform rotation calculation, the formula is shown in (2):

[0057]

[0058] ③ Convert the Cartesian coordinates of the rotated image back to image coordinates, as shown in formula (3):

[0059]

[0060] (2) Brightness Transformation

[0061] Linearly adjust the RGB channels of the image and adjust the light intensity of each channel. The brightness calculation formula is shown in (4):

[0062] g(i,j)=αf(i,j)+β (4)

[0063] Where g(i,j) is the pixel of the adjusted image, f(i,j) is the pixel of the input image, α>0, and β is the gain variable.

[0064] (3) Add Gaussian noise

[0065] In order to add interference to the data and highlight the accuracy of recognition, Gaussian noise is added to the above image. If the random variable x obeys a position parameter of μ and a variance of σ 2 The normal distribution of is, and its probability density function f(x) is shown in formula (5):

[0066]

[0067] (4) Any order combination

[0068] In the present invention, a combination of "brightness transformation + rotation", "adding Gaussian noise + rotation", "brightness transformation + adding Gaussian noise", and "brightness transformation + adding Gaussian noise + rotation" in any order is taken to increase the training data set and improve the training accuracy.

[0069] Step 3: Dataset preprocessing

[0070] Figure 7 The figure shows the result after the image is preprocessed.

[0071] (1) Piecewise linear transformation

[0072] The piecewise linear transformation function is used to enhance the grayscale areas of interest in the input image and relatively suppress those grayscale areas of no interest. The relevant formula is shown in (6):

[0073]

[0074] Among them, (0,0), (a,c), (b,d), (G max ,F max ) is the key segmentation point, g(x,y) is the gray value after transformation, is the stretching coefficient of the target area, is the transition zone retention coefficient, is the background suppression coefficient, the grayscale interval [a,b] is linearly stretched, and the grayscale intervals [0,a] and [b,F max ] is compressed.

[0075] (2) Histogram equalization method

[0076] Through a certain grayscale mapping, the input image is converted into an output image with approximately the same number of pixels at each grayscale level, so that the pixels have as many grayscale levels as possible and are as evenly distributed as possible. Assume that r and s are the grayscale levels of the original image and the image grayscale level after histogram equalization transformation, respectively. The formulas are shown in (7) and (8):

[0077]

[0078]

[0079] Where p(r) is the probability density function of r, p(s) is the probability density function of s, and T(r) is the transformation function in the histogram equalization algorithm;

[0080] Applied to discrete grayscale, assuming that the total number of pixels in an image is n, divided into L grayscale levels, the formula is shown in (9)

[0081]

[0082] Among them, s k is the image grayscale value where the kth grayscale appears, n k is the frequency of occurrence of the kth gray level, p(r k ) is the probability of the kth gray level appearing.

[0083] (3) Bilateral filtering smoothing and denoising

[0084] Since images are subject to interference from external environmental noise during digitization and transmission in reality, we use bilateral filtering to smooth and denoise the image. The formulas are shown in (10) and (11):

[0085]

[0086]

[0087] Among them, g(x,y) is the output pixel value, f(k,l) is the neighborhood pixel value, ω(i,j,k,l) ​​is the weighting coefficient of the neighborhood pixel (k,l) with (i,j) as the center point, σ d , σ r They are the filter radii of the domain kernel and range kernel respectively;

[0088] (4) Edge detection

[0089] We use the Canny operator to prioritize finding the optimal edge of the gesture:

[0090] ① Gaussian filtering: effectively removes high-frequency noise and false edges superimposed in the ideal image;

[0091] ② Calculate the gradient image and angle image, and calculate the points where the grayscale changes significantly to obtain possible edges;

[0092] ③ Perform non-maximum suppression on the gradient image, retain the points with the largest grayscale changes, and convert edges with multiple pixel widths into single-pixel edges;

[0093] ④Dual threshold segmentation, set high and low thresholds to further determine the edge position.

[0094] Step 4: Dataset Annotation

[0095] Figure 8 The figure shows the result of the dataset being labeled in Chinese.

[0096] (1) Standardize the naming of the data set and rename it in batches in the form of "######";

[0097] (2) Call the built-in Chinese font of Windows, use variable-length character encoding to realize Chinese character display, modify the font parameters, select the simhei.ttf font here, use the improved LabelImg image annotation tool to organize and annotate Chinese characters, and give specific meanings to gestures, such as 0: turn on the radio, 1: turn off the radio, 2: turn on the navigation, 3: turn off the navigation, 4: turn on the air conditioner, 5: turn off the air conditioner;

[0098] (3) Batch labeling to obtain the target dataset with the suffix .xml.

[0099] Step 5: Training Dataset

[0100] The obtained target data set is trained using the improved YOLOv5 algorithm. After the training is completed, target detection is performed on the image to be detected, and the Chinese label and recognition accuracy are displayed above the image. Fig. 9 The figure shows the detection result of the gesture to be detected.

[0101] The improved YOLOv5 algorithm structure is as follows Figure 2 As shown, the improved FPN structure is as follows Figure 3 shown.

[0102] In the original FPN structure, only three feature layers of 76*76, 38*38, and 19*19 are retained for feature fusion, and the two intermediate nodes of the 19*19 feature layer and the 76*76 feature layer are omitted. The 19*19 feature layer of the improved FPN structure is upsampled and fused with the feature layer of the intermediate dimension of 38*38. The fused output result is then connected to the 76*76 feature layer, and then a button-up bottom-up channel is added. At the same time, a jump connection of the residual edge connecting the input feature map and the output feature map is added to the feature layer with an output size of 38*38. This greatly simplifies the FPN structure and realizes efficient bidirectional cross-scale connection and multi-scale fusion, thereby improving the efficiency of the algorithm.

[0103] In terms of weights, the fast normalized fusion method is introduced. In some scenarios with large data differentiation, all data are uniformly mapped to [0,1] to reduce numerical instability. Compared with the traditional softmax method, the operation of defining weights as probabilities is changed. This method introduces nonlinear factors into neurons, improves the speed and convergence speed of the gradient descent method to find the optimal solution, and is more widely used and more efficient. Fast normalized fusion is applied to the weight of each input feature to achieve normalization effect, and the speed is greatly improved. The weight formula is shown in formula (12):

[0104]

[0105] Output is the output image result after fast normalization, w i is a learnable weight, smooth is a small amount that ensures numerical stability, i and j are the feature map indices input to the feature fusion node, and I i is the input feature map matrix, and in the experiment, smooth = 0.0001 is taken to ensure that w i ≥0, using ReLU activation function, Figure 4 The function graph of the ReLU activation function is shown, which can overcome the gradient vanishing problem, prevent overfitting, and speed up training. The output f(x) of the input vector x from the previous layer of the neural network entering the neuron after using the ReLU function is shown in (13):

[0106] f(x)=max(0,x) (13)

[0107] Fig.10 The following is a visualization result diagram related to the validation set. We performed 300 rounds of training. The five figures in the first row are the parameter values ​​during training, and the second row is the parameter values ​​when the image is to be detected. By comparing the upper and lower figures in pairs, it can be seen that the present invention can maintain good results in terms of probability measurement, classification, accuracy, and recall rate. After 20 rounds of training, the average accuracy AP of each type of image is infinitely close to 1.

[0108] Table 1 Comparison of recognition accuracy between improved YOLOv5 and traditional YOLOv5

[0109]

[0110] Experiments show that the accuracy of using the improved YOLOv5 algorithm to train and detect the gestures "0-5" to be recognized has been improved to a certain extent. The recognition rate of "0" has increased from the original 55.0% to 56.3%, the recognition rate of "1" has increased from the original 77.1% to 77.5%, the recognition rate of "2" has increased from the original 73.5% to 74.2%, the recognition rate of "3" has increased from the original 87.0% to 88.3%, the recognition rate of "4" has increased from the original 88.2% to 88.3%, and the recognition rate of "5" has increased from the original 88.6% to 89.8%.

[0111] Based on the same inventive concept, an embodiment of the present invention provides a Chinese-annotated gesture recognition system based on improved YOLOv5, comprising the following modules: an acquisition module, used to obtain a training data set, wherein the training data set includes six gestures of different people, 0 to 5; an image augmentation module, used to perform image augmentation on the obtained original data set, including rotation affine transformation, brightness transformation, Gaussian noise addition operation, and arbitrary order combination operation; an image preprocessing module, used to perform preprocessing operation on the obtained data set, using piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection operation to obtain a target image, and the target image constitutes a target data set; a data set annotation module, used to annotate the preprocessed target data set using a modified LabelImg annotation tool, add specific command meanings to each gesture of 0 to 5, and obtain a target data set with .xml as the suffix; a model training module, used to improve the FPN structure of the YOLOv5 algorithm, introduce a fast normalization fusion strategy, and use the improved algorithm to train the obtained target data set; and a gesture recognition module, used to perform target detection on the image to be detected after training, and display the Chinese label and recognition accuracy above the image. The specific implementation details of each module refer to the above method embodiment and will not be repeated here.

[0112] Based on the same inventive concept, the present invention provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the Chinese annotation gesture recognition method based on the improved YOLOv5 is implemented.

[0113] The above embodiments are only used to help understand the method and core idea of ​​the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A Chinese annotation gesture recognition method based on improved YOLOv5, characterized in that: The steps include: Step 1: Select a training data set, which includes 0 to 5 gestures of different people; Step 2: Augment the original data set by image augmentation, including rotation affine transformation, brightness transformation, Gaussian noise addition, and any combination of operations in any order; Step 3: Perform preprocessing operations on the obtained data set, using piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection operations to obtain the target image, and the target image constitutes the target data set; Step 4: Use the modified LabelImg annotation tool to annotate the preprocessed target dataset, add specific command meanings to each gesture from 0 to 5, and obtain the target dataset with the suffix .xml; Step 5: Improve the FPN structure of the YOLOv5 algorithm, introduce a fast normalization fusion strategy, and use the improved algorithm to train the obtained target data set; Step 6: After training is completed, target detection is performed on the image to be detected, and the Chinese label and recognition accuracy are displayed above the image; The FPN structure of the improved YOLOv5 algorithm is as follows: only three feature layers of 76*76, 38*38, and 19*19 are retained on the original FPN structure for feature fusion, and two intermediate nodes of the 19*19 feature layer and the 76*76 feature layer are omitted, the 19*19 feature layer of the improved FPN structure is upsampled, and fused with the feature layer of the intermediate dimension of 38*38, and then the fused output result is connected with the 76*76 feature layer, and then a button-up bottom-up channel is added, and at the same time, a skip connection of the residual edge connecting the input feature map and the output feature map is added to the feature layer with an output size of 38*38, so as to simplify the FPN structure and realize bidirectional cross-scale connection and multi-scale fusion; The fast normalization fusion strategy is: compared with the traditional softmax method, it changes the operation of defining weights as probabilities, introduces nonlinear factors into neurons, and improves the speed and convergence speed of the gradient descent method to find the optimal solution. The weight formula is as follows: Output is the output image result after fast normalization, w i is a learnable weight, smooth is a small amount that ensures numerical stability, i and j are the feature map indices input to the feature fusion node, and I i is the input feature map matrix. In order to ensure that w i ≥0, where the ReLU activation function is used.

2. The Chinese annotation gesture recognition method based on improved YOLOv5 according to claim 1 is characterized in that: In the step 4, the Windows version of LabelImg image annotation tool is modified, the character type is changed, the Windows internal Chinese font is called, and variable-length character encoding is used to realize Chinese character display, so that the open file path can contain Chinese directories, and the functions of annotating labels in Chinese and displaying Chinese labels can also be realized.

3. The Chinese annotation gesture recognition method based on improved YOLOv5 according to claim 1 is characterized in that: In the step 2, the rotated image is filled with black.

4. The Chinese annotation gesture recognition system based on improved YOLOv5 is characterized by: Includes the following modules: An acquisition module is used to obtain a training data set, which includes six gestures (0 to 5) of different people; The image augmentation module is used to augment the original data set, including rotation affine transformation, brightness transformation, adding Gaussian noise operation, and any combination of operations in any order; The image preprocessing module is used to perform preprocessing operations on the obtained data set, using piecewise linear transformation, histogram equalization method, smoothing denoising and edge detection operations to obtain the target image, which constitutes the target data set; The data set annotation module is used to annotate the preprocessed target data set using the modified LabelImg annotation tool, adding specific command meanings to each gesture from 0 to 5, and obtaining the target data set with the suffix .xml; The model training module is used to improve the FPN structure of the YOLOv5 algorithm and introduce a fast normalization fusion strategy, and use the improved algorithm to train the obtained target data set; And, the gesture recognition module is used to perform target detection on the image to be detected after training is completed, and display the Chinese label and recognition accuracy above the image; The FPN structure of the improved YOLOv5 algorithm is as follows: only three feature layers of 76*76, 38*38, and 19*19 are retained on the original FPN structure for feature fusion, and two intermediate nodes of the 19*19 feature layer and the 76*76 feature layer are omitted, the 19*19 feature layer of the improved FPN structure is upsampled, and fused with the feature layer of the intermediate dimension of 38*38, and then the fused output result is connected with the 76*76 feature layer, and then a button-up bottom-up channel is added, and at the same time, a skip connection of the residual edge connecting the input feature map and the output feature map is added to the feature layer with an output size of 38*38, so as to simplify the FPN structure and realize bidirectional cross-scale connection and multi-scale fusion; The fast normalization fusion strategy is: compared with the traditional softmax method, it changes the operation of defining weights as probabilities, introduces nonlinear factors into neurons, and improves the speed and convergence speed of the gradient descent method to find the optimal solution. The weight formula is as follows: Output is the output image result after fast normalization, w i is a learnable weight, smooth is a small amount that ensures numerical stability, i and j are the feature map indices input to the feature fusion node, and I i is the input feature map matrix. In order to ensure that w i ≥0, where the ReLU activation function is used.

5. The Chinese annotation gesture recognition system based on improved YOLOv5 according to claim 4 is characterized in that: In the data set annotation module, the Windows version LabelImg image annotation tool is modified to change the character type, call the Windows internal Chinese font, and use variable-length character encoding to realize Chinese character display, so that the open file path can contain Chinese directories, and the functions of annotating labels in Chinese and displaying Chinese labels can also be realized.

6. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the Chinese annotation gesture recognition method based on the improved YOLOv5 is implemented according to any one of claims 1 to 3.