Respirator waveform classification and identification method based on improved YOLO target detection

Abstract

The invention discloses a breathing machine waveform classification and identification method based on improved YOLO target detection, and the method comprises the steps: 1, obtaining a normal breathing machine waveform image and an abnormal breathing machine waveform image as a training sample, and enabling the training sample to represent the corresponding relation between the shapes and parameters of different breathing machine waveforms and the mechanical ventilation states of the different breathing machine waveforms; step 2, performing batch labeling on the obtained training samples, obtaining mechanical ventilation state text information corresponding to the breathing machine waveform image and position information corresponding to the breathing machine waveform in the training samples, and establishing a breathing machine waveform recognition library; 3, introducing an improved preprocessing module, processing a training sample by using an improved YOLO target detection algorithm training model, and carrying out breathing machine waveform image classification and recognition training to obtain a breathing machine waveform image classification and recognition model; and step 4, completing automatic classification and identification of the breathing machine waveform image by using the breathing machine waveform image classification and identification model. According to the method, the waveform of the respirator can be detected online in real time, and the abnormal waveform classification accuracy is high.

Description

Technical Field

[0001] The present invention relates to the technical field of ventilators, and in particular to a ventilator waveform classification and recognition method based on improved YOLO target detection. Background Art

[0002] A ventilator is a medical device used for mechanical ventilation. It is often used in the treatment of patients with dyspnea and respiratory failure and has important applications in the field of critical care medicine. With the widespread application of ventilators, the mortality rate of patients with dyspnea and respiratory failure has dropped significantly. In modern times, with the rapid development of electronic technology, ventilators have evolved from primitive mechanical ventilation equipment to today's complex medical equipment that integrates mechanical ventilation and the collection and display of key physiological parameters of patients.

[0003] The ventilator waveform is a waveform drawn by the ventilator by collecting various parameters of the patient during mechanical ventilation. It is one of the important data provided by the ventilator. According to the shape and parameters of the ventilator waveform, medical staff can judge the current state of mechanical ventilation and find possible problems between the breathing circuit formed by the ventilator and the patient. If the problem is not located and solved in time, it is easy to cause the mechanical ventilation effect to deteriorate and even cause lung damage and pneumonia caused by mechanical ventilation. However, medical staff have been in an environment with high work intensity and high work pressure for a long time. It is difficult to judge the effect of the current mechanical ventilation and whether there are abnormal conditions in time through the ventilator waveform. Therefore, it is of great significance to detect mechanical ventilation abnormalities and evaluate the effect of mechanical ventilation through automatic detection and classification of ventilator waveforms to maintain the stability of mechanical ventilation and reduce the work pressure of medical staff.

[0004] With the development of computer vision, machine learning and other technologies, related technologies have been widely used in various target detection fields and have achieved good performance in various target detection scenarios. However, traditional computer vision processing methods, such as image segmentation and pattern matching algorithms based on color, grayscale, preset features, etc., perform poorly in the face of complex target detection tasks. In actual scenarios, these methods are greatly affected by external factors such as image shooting angle and brightness, and the accuracy is not ideal. The ventilator waveform has the characteristics of diverse shapes, numerous classifications, and small morphological differences between different classifications. Combined with actual application scenarios, the captured ventilator waveform images may also have problems such as tilt and high light pollution. Various factors show that it is difficult to complete the classification and detection tasks of ventilator waveform images using traditional computer vision and image processing methods. The application of deep learning technology in computer vision has successfully solved a large number of difficult target detection tasks that are difficult to complete with traditional methods. YOLO, as a deep learning target detection model, has been successfully applied in a large number of fields. The application of the YOLO target detection model to the classification and recognition tasks of ventilator waveform images has important application value.

[0005] Therefore, technicians in this field are committed to developing a ventilator waveform classification and recognition method based on improved YOLO target detection. Summary of the invention

[0006] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to realize the universal and easy real-time export of data exported by ventilators of different models and the online detection of ventilator waveforms.

[0007] To achieve the above object, the present invention provides a ventilator waveform classification and recognition method based on improved YOLO target detection, comprising the following steps:

[0008] Step 1, obtaining normal ventilator waveforms and abnormal ventilator waveform images as training samples, wherein the training samples characterize the corresponding relationship between shapes and parameters of different ventilator waveforms and their mechanical ventilation states;

[0009] Step 2: batch label the acquired training samples, obtain the mechanical ventilation status text information corresponding to the ventilator waveform image and the position information of the corresponding ventilator waveform in the training sample, and establish a ventilator waveform recognition library;

[0010] Step 3, introducing an improved preprocessing module, using the improved YOLO target detection algorithm training model to process training samples, performing ventilator waveform image classification and recognition training, and obtaining a ventilator waveform image classification and recognition model;

[0011] Step 4: Use the ventilator waveform image classification and recognition model to complete the automatic classification and recognition of the ventilator waveform image.

[0012] Furthermore, the training samples include normal ventilator waveforms and abnormal ventilator waveform images obtained by photographing a waveform display screen of a ventilator in normal operation having the function of displaying one or more of the three waveforms, namely, pressure-time waveform, flow rate-time waveform and volume-time waveform. The training samples are of arbitrary size and are 3-channel RGB color images.

[0013] Furthermore, the ventilator waveform image recognition library is composed of an image library and a label library. The image library stores training sample images, and the training sample images are uniformly scaled to a size of 640*640, and data augmentation is performed by changing image brightness parameters, random scaling, random cropping and mirror flipping methods. The label library stores position information of different ventilator waveforms in each training sample and category information of different ventilator waveforms in the training sample to form corresponding training sample labels. At the same time, the label library also stores corresponding information between labels and images in the image library.

[0014] Furthermore, the recognition library is divided into a training set, a test set and a validation set in a ratio of 8:1:1. After loading the pre-training weights, the YOLO network is trained using training samples. When the loss function value no longer decreases significantly, the training process is terminated and the model weights at this time are saved to obtain a ventilator waveform image classification and recognition model.

[0015] Furthermore, the improved YOLO target detection algorithm is YOLOv5, and the improved preprocessing module is introduced into the Input part, and the improved preprocessing module includes the following steps:

[0016] Step (1) Use the YOLO model to separate the area where the ventilator waveform is located from the image.

[0017] The YOLO small model adopts the YOLOv5s structure, uses the same ventilator waveform image recognition library for training, uses a simplified label form, no longer distinguishes the mechanical ventilation state corresponding to the ventilator waveform, and functions to directly extract the area where the ventilator waveform is located from the complex background;

[0018] Step (2) using Huffman line detection to perform tilt correction on the image of the area where the waveform is located;

[0019] Step (3) uses color threshold segmentation to further modify the area where the waveform is located.

[0020] Furthermore, the step (2) comprises the following steps:

[0021] Step 1) Perform grayscale transformation on the image, where the grayscale transformation formula is:

[0022] g=0.299R+0.587G+0.114B

[0023] Where g is the grayscale value obtained by grayscale transformation, R, G and B are the pixel values ​​corresponding to the red, green and blue channels respectively;

[0024] Step 2) performing Canny edge detection on the grayscale image after grayscale transformation;

[0025] Step 3) Perform Huffman line detection on the edge detection result;

[0026] Step 4) Calculate the affine matrix based on the line detection results to complete the image rotation.

[0027] Furthermore, in step 3), for the results of Huffman line detection, the results are sorted according to the straight line inclination angles, straight lines with inclination angles greater than a certain angle are filtered out, the inclination angles of the remaining straight lines are averaged, and the average value is used as the Huffman line detection result.

[0028] Furthermore, the step (3) comprises the following steps:

[0029] Step 1) Convert the original image from RCG color space to HSV color space;

[0030] Step 2) Perform color segmentation on the image according to a preset color threshold group to obtain a mask;

[0031] Step 3) performing contour detection suppression on the mask according to the set threshold;

[0032] Step 4) Scan line by line to determine the upper and lower bounds of the cropped image based on the suppressed mask.

[0033] Furthermore, the classification and recognition training includes the following steps:

[0034] Step (1) Establish a YOLO target detection training model based on Darknet-53, and the output size of the model is 19*19*39, 38*38*39 and 76*76*39;

[0035] Step (2) Determine the size of the initial detection box using the k-means algorithm based on the morphology and distribution characteristics of the training samples;

[0036] Step (3) Based on the initial detection frame, forward propagation generates a predicted frame, compares the predicted frame with the real frame in the training sample, calculates the difference between the two, and back propagates to update the weights in the model;

[0037] Step (4) takes the predicted box as the new detection box, alternately performs forward propagation and backward propagation, and iterates the model weight parameters;

[0038] Step (5) When the training reaches the preset IoU and Loss values, the training process is stopped, the weight parameters of the model are saved, and the ventilator waveform image classification and recognition model is obtained. IoU represents the intersection over union ratio, and Loss is the deviation between the actual output value of the model and the sample label value.

[0039] Furthermore, the automatic classification and recognition of the ventilator waveform image using the ventilator waveform image classification and recognition model includes the following steps:

[0040] Step (1) inputting the ventilator waveform image into the trained ventilator waveform image classification and recognition model for prediction, and obtaining the output prediction box confidence, classification confidence and prediction box position according to the prediction value;

[0041] Step (2) sorting the classification confidences and selecting the classification with the largest confidence as the predicted classification;

[0042] Step (3) performs non-maximum suppression on the prediction results, and the remaining prediction results after suppression are used as the final classification recognition results.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] (1) It can realize real-time online detection of ventilator waveform, independent of the specific ventilator model.

[0045] (2) Under various lighting conditions and data acquisition device placement conditions, the accuracy of abnormal waveform classification can be maintained at a high level.

[0046] (3) It can not only detect common abnormalities such as air leakage and water accumulation in the breathing circuit, but also detect endogenous positive end-expiratory pressure. In addition, new functions can be flexibly added according to actual conditions to meet more needs.

[0047] (4) High versatility, low coupling, good maintainability, and easy expansion and integration.

[0048] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of an embodiment of the present invention;

[0050] Figure 2 It is a structural diagram of a ventilator waveform image classification and recognition model of the present invention;

[0051] Figure 3 This is a flow chart of the present invention regarding the introduction of a pre-processing module;

[0052] Figure 4 This is a sub-flow chart of the tilt correction introduced into the pre-processing module of the present invention;

[0053] Figure 5 The present invention is about introducing a further correction sub-flow chart into the pre-processing module;

[0054] Figure 6 The invention is about introducing a sub-flow chart for automatic classification and recognition of ventilator waveform images. DETAILED DESCRIPTION

[0055] The following describes the preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0056] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.

[0057] like Figure 1 As shown, the present invention provides a ventilator waveform image classification and recognition method based on YOLO target detection, comprising the following steps:

[0058] Step S101, acquiring images including normal ventilator waveforms and abnormal ventilator waveforms as training samples, wherein the training samples represent the corresponding relationship between shapes and parameters of different ventilator waveforms and their mechanical ventilation states;

[0059] Step S102, batch labeling the acquired training samples, obtaining the mechanical ventilation status text information corresponding to the ventilator waveform image and the position information of the corresponding ventilator waveform in the training sample, and establishing a ventilator waveform recognition library;

[0060] Step S103, introducing an improved preprocessing module, using an improved YOLO target detection algorithm training model to process training samples, performing ventilator waveform image classification and recognition training, and obtaining a ventilator waveform image classification and recognition model;

[0061] Step S104, using the ventilator waveform image classification and recognition model to complete automatic classification and recognition of the ventilator waveform image.

[0062] In step S101, the images including normal ventilator waveforms and abnormal ventilator waveforms captured by the camera are obtained by capturing the waveform display screen of a ventilator in normal operation having the function of displaying one or more of the three waveforms including pressure-time waveform, flow rate-time waveform and volume-time waveform, and the captured results are exported as 3-channel RGB color images with different sizes.

[0063] In step S102, the annotation tool CVAT is used to annotate the color image exported in step S1, select the location of the ventilator waveform in the image, and mark the mechanical ventilation state corresponding to the ventilator waveform. After completing the annotation of all pictures, the annotation information is exported to the YOLO annotation format to obtain the label file corresponding to the ventilator waveform image. The label file contains the mechanical ventilation state corresponding to the waveform in the image and the real frame coordinates obtained by the frame selection. The image is uniformly scaled to 640*640 size, and data augmentation is performed by changing parameters such as image brightness, random scaling, random cropping, mirror flipping, and other methods. Finally, a ventilator waveform image recognition library consisting of an image library and a label library is obtained.

[0064] like Figure 2 As shown, the YOLO target detection algorithm used in step S103 is YOLOv5. YOLOv5 consists of four parts, namely Input, Backbone, Neck and Predication. Among them, the input part is the data to be input into the YOLO network, which is a 3-channel 640*640RGB color picture. Backbone adopts the CSPDarknet-53 structure, which consists of 52 convolutional layers and 1 fully connected layer to generate feature maps. The Neck part is the SPPF and CSP-PAN structure, which is used to generate feature pyramids, suitable for the same target of different sizes. The Predication part convolves the feature map to obtain output feature maps with scales of 19*19*39, 38*38*39 and 76*76*39 respectively. Before training, the k-means algorithm is used to perform cluster analysis on the data in the training samples, and three prior boxes of sizes 10*13, 16*30 and 33*23 are selected as the prior boxes used for the output feature map of 76*76*39; the prior boxes of sizes 30*61, 62*45 and 59*119 are assigned to the output feature map of 38*38*39; the prior boxes of sizes 116*90, 156*198 and 373*326 are assigned to the output feature map of 19*19*39.

[0065] The loss function used in the training model is CIoU_Loss, and its function expression is:

[0066]

[0067] Where IoU is the intersection over union ratio. Let the size of the intersection of the two rectangular boxes be A and the size of the area union be B, then:

[0068]

[0069] Dc is the Euclidean distance between the center points of the two rectangular boxes, Dr is the diagonal distance between the circumscribed rectangles of the two rectangular boxes, and v is the consistency parameter, which is calculated as follows:

[0070]

[0071] Where w and h represent the width and height of the rectangular box respectively, the subscript t represents the real box, and p represents the predicted box.

[0072] The recognition library obtained in step S102 is divided into a training set, a test set and a validation set in a ratio of 8:1:1. After loading the pre-training weights, the YOLO network is trained using the training samples. When the loss function value no longer decreases significantly, the training process is terminated and the model weights at this time are saved to obtain a ventilator waveform image classification and recognition model.

[0073] like Figure 3 As shown, in step S103, the improved preprocessing module is introduced into the Input part, and the introduced improved preprocessing module specifically includes the following steps:

[0074] Step S103.1, using the YOLO model to separate the area where the ventilator waveform is located from the image;

[0075] Step S103.2, using Huffman line detection to perform tilt correction on the image of the area where the waveform is located;

[0076] Step S103.3, further correct the area where the waveform is located by using color threshold segmentation.

[0077] The YOLO small model adopts the YOLOv5s structure and is trained using the same ventilator waveform image recognition library. It uses a streamlined label format and no longer distinguishes the mechanical ventilation state corresponding to the ventilator waveform. Its function is to directly extract the area where the ventilator waveform is located from the complex background for subsequent processing.

[0078] like Figure 4 As shown, step S103.2 includes the following steps:

[0079] Step S103.2.1, performing grayscale transformation on the image;

[0080] Step S103.2.2, performing Canny edge detection on the grayscale image after grayscale transformation;

[0081] Step S103.2.3, performing Huffman line detection on the edge detection result;

[0082] Step S103.2.4, calculate the affine matrix according to the straight line detection result to complete the image rotation.

[0083] The grayscale transformation in step S103.2.1 adopts the following formula:

[0084] g=0.299R+0.587G+0.114B

[0085] Where g is the grayscale value obtained by grayscale transformation, and R, G, and B are the pixel values ​​corresponding to the red, green, and blue channels, respectively.

[0086] In step S103.2.3, the results of the Huffman line detection are sorted according to the line inclination angles, and the lines with inclination angles greater than a certain angle are filtered out. The inclination angles of the remaining lines are averaged and the average value is used as the Huffman line detection result.

[0087] like Figure 5 As shown, step S103.3 includes the following steps:

[0088] Step S103.3.1, converting the original image from the RCG color space to the HSV color space;

[0089] Step S103.3.2, performing color segmentation on the image according to a preset color threshold group to obtain a mask;

[0090] Step S103.3.3, performing contour detection suppression on the mask according to the set threshold;

[0091] Step S103.3.4, based on the suppressed mask, scan line by line to determine the upper and lower bounds of the cropped image.

[0092] like Figure 6 As shown, in step S104, the automatic classification and recognition of the ventilator waveform image using the ventilator waveform image classification and recognition model includes the following steps:

[0093] Step S104.1, inputting the ventilator waveform image into the trained ventilator waveform image classification and recognition model for prediction, and obtaining the prediction box confidence, classification confidence and prediction box position;

[0094] Step S104.2, sorting the classification confidences, and selecting the classification with the highest confidence as the predicted classification;

[0095] Step S104.3, using the NMS method to perform maximum value suppression, and taking the remaining result after suppression as the final classification and recognition result.

[0096] The present invention can automatically classify and identify ventilator waveform images, determine the current mechanical ventilation status, and solve the problems of medical staff not being able to use ventilator waveforms to determine the mechanical ventilation status in a timely manner or making misjudgments. It can cope with various complex environments, and can still maintain the accuracy of ventilator waveform image classification when there are tilt problems and high light pollution in the image. It does not damage the existing ventilator hardware equipment and has a wide range of applicability and economic value.

[0097] The present invention has the technical advantages of strong versatility, low coupling, good maintainability, and easy expansion and integration. In terms of versatility, the automatic classification and detection of ventilator waveforms is performed based on image recognition, which is independent of specific ventilator models and waveform display forms. It can cope with the characteristics of different ventilator brands and different waveform display settings in actual application scenarios. In terms of coupling, the data acquisition and analysis modules used are decoupled from the ventilator and can be used without modifying the existing ventilator equipment, which greatly reduces the application cost and difficulty. In terms of maintainability, the hardware equipment used is relatively simple, and the algorithm used can be hot-updated without interrupting the existing detection process. The hardware modules used in the method have mature products available for replacement and selection, and have good maintainability. In terms of scalability, rapid algorithm expansion and iteration can be achieved based on changes in scenarios and changes in demand, and the output results of the method can be used for further integration and storage with other equipment and software for subsequent clinical analysis.

[0098] The main performance indicators of the present invention are the processing time per frame and the recognition accuracy. The former can be measured by the number of frames per second (FPS), and the latter is measured by percentage. In actual production applications, the change rate of the ventilator waveform usually does not exceed 24FPS, which is much lower than the time taken by the method to process each frame on a better hardware platform. The overall recognition accuracy of the method is above 95%, which has a good recognition accuracy.

[0099] In terms of production implementation, the present invention can be quickly deployed in clinical application scenarios according to actual conditions to assist doctors in judging the patient's respiratory status and collecting data. The method can also be quickly iterated and modified according to actual conditions and changes in demand, and has high usability.

[0100] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A ventilator waveform classification and recognition method based on improved YOLO target detection, characterized in that: The following steps are involved: Step 1, obtaining normal ventilator waveforms and abnormal ventilator waveform images as training samples, wherein the training samples characterize the corresponding relationship between shapes and parameters of different ventilator waveforms and their mechanical ventilation states; Step 2: batch label the acquired training samples, obtain the mechanical ventilation status text information corresponding to the ventilator waveform image and the position information of the corresponding ventilator waveform in the training sample, and establish a ventilator waveform recognition library; Step 3, introducing an improved preprocessing module, using the improved YOLO target detection algorithm training model to process training samples, performing ventilator waveform image classification and recognition training, and obtaining a ventilator waveform image classification and recognition model; Step 4: Use the ventilator waveform image classification and recognition model to complete the automatic classification and recognition of the ventilator waveform image.

2. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The training samples include normal ventilator waveforms and abnormal ventilator waveform images from the waveform display screen of a ventilator in normal operation, which has the function of displaying one or more of the three waveforms including pressure-time waveform, flow rate-time waveform and volume-time waveform. The training samples are of arbitrary size and are 3-channel RGB color pictures.

3. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The ventilator waveform image recognition library is composed of an image library and a label library. The image library stores training sample images, which are uniformly scaled to a size of 640*640, and data augmentation is performed by changing image brightness parameters, random scaling, random cropping and mirror flipping methods. The label library stores position information of different ventilator waveforms in each training sample and category information of different ventilator waveforms in the training sample to form corresponding training sample labels. At the same time, the label library also stores corresponding information between labels and images in the image library.

4. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The recognition library is divided into a training set, a test set and a validation set in a ratio of 8:1:

1. After loading the pre-training weights, the YOLO network is trained using the training samples. When the loss function value no longer decreases significantly, the training process is terminated and the model weights at this time are saved to obtain a ventilator waveform image classification and recognition model.

5. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The improved YOLO target detection algorithm is YOLOv5, and the improved preprocessing module is introduced into the Input part. The improved preprocessing module includes the following steps: Step (1) Use the YOLO model to separate the area where the ventilator waveform is located from the image. The YOLO small model adopts the YOLOv5s structure, uses the same ventilator waveform image recognition library for training, uses a simplified label form, no longer distinguishes the mechanical ventilation state corresponding to the ventilator waveform, and functions to directly extract the area where the ventilator waveform is located from the complex background; Step (2) using Huffman line detection to perform tilt correction on the image of the area where the waveform is located; Step (3) uses color threshold segmentation to further modify the area where the waveform is located.

6. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 5, characterized in that, The step (2) comprises the following steps: Step 1) Perform grayscale transformation on the image, where the grayscale transformation formula is: g=0.299R+0.587G+0.114B Where g is the grayscale value obtained by grayscale transformation, R, G and B are the pixel values ​​corresponding to the red, green and blue channels respectively; Step 2) performing Canny edge detection on the grayscale image after grayscale transformation; Step 3) Perform Huffman line detection on the edge detection result; Step 4) Calculate the affine matrix based on the line detection results to complete the image rotation.

7. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 6, characterized in that, In the step 3), the results of the Huffman line detection are sorted according to the straight line inclination angles, the straight lines with inclination angles greater than a certain angle are filtered out, the inclination angles of the remaining straight lines are averaged, and the average value is used as the Huffman line detection result.

8. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 5, characterized in that, The step (3) comprises the following steps: Step 1) Convert the original image from RCG color space to HSV color space; Step 2) Perform color segmentation on the image according to a preset color threshold group to obtain a mask; Step 3) performing contour detection suppression on the mask according to the set threshold; Step 4) Scan line by line to determine the upper and lower bounds of the cropped image based on the suppressed mask.

9. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The classification and recognition training comprises the following steps: Step (1) Establish a YOLO target detection training model based on Darknet-53, and the output size of the model is 19*19*39, 38*38*39 and 76*76*39; Step (2) Determine the size of the initial detection box using the k-means algorithm based on the morphology and distribution characteristics of the training samples; Step (3) Based on the initial detection frame, forward propagation generates a predicted frame, compares the predicted frame with the real frame in the training sample, calculates the difference between the two, and back propagates to update the weights in the model; Step (4) takes the predicted box as the new detection box, alternately performs forward propagation and backward propagation, and iterates the model weight parameters; Step (5) When the training reaches the preset IoU and Loss values, the training process is stopped, the weight parameters of the model are saved, and the ventilator waveform image classification and recognition model is obtained, where IoU represents the intersection over union ratio and Loss is the deviation between the actual output value of the model and the sample label value.

10. The ventilator waveform classification and recognition method based on improved YOLO target detection as claimed in claim 1, characterized in that, The automatic classification and recognition of ventilator waveform images using the ventilator waveform image classification and recognition model includes the following steps: Step (1) inputting the ventilator waveform image into the trained ventilator waveform image classification and recognition model for prediction, and obtaining the output prediction box confidence, classification confidence and prediction box position according to the prediction value; Step (2) sorting the classification confidences and selecting the classification with the highest confidence as the predicted classification; Step (3) performs non-maximum suppression on the prediction results, and the remaining prediction results after suppression are used as the final classification recognition results.

Images