Method and system for recognizing specific region color of safety helmet based on deep learning

Abstract

A safety helmet specific area color identification method based on deep learning, by identifying the color of the screen display in the center of the brim of the safety helmet, the identity of the staff is obtained, and whether it has the right of passage is judged. The method comprises the following steps: first, collecting image information to obtain the moving human body target in the image. Second, input the target into the improved YOLOv5 target detection model, use the soft and strong attention mechanism to determine the position of the safety helmet and the screen in the center of the brim, and judge whether the human body target correctly wears the safety helmet through the position information. Then, the multi-target detection result is input into the ResNet and classification network for color identification, and the SENet and priority module are added to weight the color information of the specific area; finally, the relevant information is integrated. The safety helmet specific area color identification and judgment of the application can identify the right of passage of different personnel, and can be used to judge the right of passage of the staff.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, and in particular to a safety helmet specific area color recognition method and system based on deep learning. BACKGROUND

[0002] Safety production is the most core requirement in industrial production process. In industrial production, different workers need to supervise, detect, maintain and repair different industrial procedures. In the face of high-risk equipment and different regional division of labor, it is necessary for different workers to wear safety helmets when entering different sites. In industrial sites, it is required to monitor whether the production mode in the production process meets the requirements and whether the workers can work in their corresponding work areas in real time. In addition, there are many key protection areas in chemical plant areas, and there are strict requirements for the number and duration of personnel entering, so personnel can be planned in important areas of the plant, and the safety level of personnel wearing safety helmets of different colors in specific areas can be identified to realize functions such as "boundary alarm" and "personnel identification".

[0003] Traditional safety helmet color recognition mainly aims at personnel identity recognition and whether the safety helmet is correctly worn to make a judgment. The traditional computer vision is used to recognize the color of the whole safety helmet. After the safety helmet target detection model detects the safety helmet area, the traditional computer vision method is used to analyze and judge the color of the safety helmet. The disadvantage of this method is that in dark light, the safety helmet color cannot be accurately recognized, and the specific area color cannot be effectively recognized, which has certain interference to the specific area color recognition. SUMMARY

[0004] In view of the above technical problems that the safety helmet and the specific area color of the safety helmet cannot be clearly recognized in dark light, the present technical solution provides a safety helmet specific area color recognition method and system based on deep learning. A strong attention SENet module is added to the ResNet network to realize strong attention to the specific area color on the safety helmet, and the recognition performance of the specific color area is improved. The above problems can be effectively solved.

[0005] The present application is realized by the following technical solutions:

[0006] A safety helmet specific area color recognition method based on deep learning, and the specific recognition method includes the following steps:

[0007] Step 1: Obtain a video monitoring picture, and continuously acquire video frames of the obtained video;

[0008] Step 2: Collect image information, analyze the obtained video frames, and when a human target is detected in the image information to be tested, intercept the area image of the position of the human target;

[0009] Step 3: Perform human target and background segmentation processing on the human region image, and uniformly transform the background pixels into preset background color pixels to obtain a human target image;

[0010] Step 4: Input the human target region into a pre-established convolutional neural network model, and perform multi-target parallel detection on the image information according to a pre-set target classification; the specific operation mode is as follows: input the human target region image into a human detection model, which is a target detection model based on YOLOv5, and can identify the overall safety helmet and the position of the specific area on the safety helmet worn by the human body;

[0011] The overall safety helmet and the position of the specific area on the safety helmet extracted by the target detection model of YOLOv5 can be basically determined, but for subsequent feature extraction of the specific area on the safety helmet, further spatial transformation is needed for the currently extracted overall safety helmet and the position of the specific area on the safety helmet, and a spatial transformer module needs to be added in the YOLOv5 feature extraction network to perform corresponding spatial transformation on the spatial domain information of the position of the safety helmet and the specific area on the safety helmet in the image, so as to extract the key information;

[0012] Step 5: According to the parallel detection result, the detection result at least includes the information of the overall safety helmet color and the specific area position and color on the safety helmet.

[0013] Further, the target detection model based on YOLOv5 is further improved, and the specific improvement mode is as follows: an attention mechanism is applied in the target detection model of YOLOv5, and strong attention and soft attention are used to distinguish the attention of the position of the specific area on the safety helmet and the overall safety helmet, respectively; specifically, a strong attention SE Net module is added behind a high-level convolution layer or a pooling layer of a Backbone network of YOLOv5, and a soft attention FPN module is added behind a high-level convolution layer or a pooling layer of a Neck network of YOLOv5, so as to effectively improve the accuracy and robustness of the model.

[0014] Further, the acquisition of the information of the overall safety helmet color and the specific area color on the safety helmet includes the following specific steps:

[0015] After obtaining the human target region, the overall safety helmet color and the specific area color on the safety helmet need to be further identified and classified;

[0016] According to the improved YOLOv5-based target detection model, the accurate positions of the safety helmet and the specific area on the safety helmet are obtained, and the output results are used as the input of the ResNet network model to extract the features of the safety helmet and the specific area on the safety helmet. Finally, the extracted features are input into the classification network for color classification.

[0017] A SENet module is added after the third Block of the ResNet-50 in the ResNet network model as a strong attention mechanism. The main purpose of the SENet module is to enhance the network's attention to specific features and improve network performance. Specifically, the position features of the safety helmet and the specific area on the safety helmet output by YOLOv5 are introduced into the SENet module, and the Squeeze-and-Excitation operation is used to enhance the attention to the safety helmet color.

[0018] The two features are combined together by using a feature fusion method to form a unified feature representation. Through feature fusion, information from the target detection model and the specific area recognition module is combined to obtain a more comprehensive and informative feature representation.

[0019] Further, the extraction of color information of the safety helmet and the specific area of the safety helmet includes the following specific steps:

[0020] An auxiliary network is added before the strong attention module of the ResNet model to predict the position information of the specific area of the safety helmet, and the prediction results are provided to the strong attention module. Specifically, an additional output is added in YOLOv5, which will predict the display screen position of the safety helmet before the strong attention module. The strong attention module can use the specific area as the focus point to further identify the color on the safety helmet after positioning the specific area.

[0021] The priority module is used to perform weighted average algorithm on the color of the safety helmet and the specific area on the safety helmet to adjust the weight of the specific area, so that even if the color of the safety helmet cannot be accurately identified, the information can be judged by identifying the color of the specific area.

[0022] Further, the information of the overall color of the safety helmet and the specific area color on the safety helmet, the specific color requirements and the right to pass are as follows:

[0023] (I) Basic color

[0024] Chemical industry safety helmet color general provisions:

[0025] 1) White: project leaders and workshop team management personnel;

[0026] 2) Blue: Special personnel and mechanical operation maintenance personnel;

[0027] 3) Yellow: Construction team personnel;

[0028] 4) Red: Project full-time safety officer;

[0029] In order not to destroy the safety helmet color specified in the chemical industry, under the premise that the main color of the safety helmet does not change, a specific area is selected, a display screen is installed, the color of the display screen is changed, and the activity area of different personnel and the identity of the related personnel are distinguished;

[0030] (II) Regional access color

[0031] 1) Highest access right: green; The specific area is green, which is matched with any of the above four theme colors, and the safety helmet with the highest access right can access the whole area;

[0032] 2) Medium access right: purple; The specific area is purple, which is matched with any of the above four theme colors, and the safety helmet with medium access right can access the medium and low level areas;

[0033] 3) Lowest access right: orange; The specific area is orange, which is matched with any of the above four theme colors, and the safety helmet with the lowest access right can only access the low level area.

[0034] Further, the specific operation manner of the human body target and background segmentation processing in step 3 to obtain the human body target frame image and distinguish the human body target and background pixels is: according to the collected image information, continuous frame images are obtained, so that the difference value of each pixel point in the current frame image and the previous frame image is obtained, the difference value is compared with the preset threshold value, when the difference value exceeds the preset threshold value range, the corresponding pixel point is regarded as a foreground image, otherwise, it is regarded as a background image, and the moving object image in the image information is obtained through the foreground image, that is, the human body target.

[0035] A safety helmet specific area color recognition system based on deep learning can be applied to the safety helmet specific area color recognition method based on deep learning, and the recognition system comprises:

[0036] An image processing module acquires image information and moving object images in the image information, and separates human body targets and backgrounds;

[0037] A position recognition module is used for multi-target parallel detection of input to YOLOv5 to obtain safety helmet and display screen positions, and the detection result at least includes safety helmet overall color and safety helmet specific area position information;

[0038] A position determination module is configured to determine whether a safety helmet is worn by a person in the collected image information according to the parallel detection result, and if the position of the safety helmet is not recognized, it is determined that the person does not wear the safety helmet, and an alarm is given.

[0039] A color identification module is configured to identify the color of the input to the ResNet and the classification network, and perform classification.

[0040] A color determination module is configured to determine whether the access level of the staff member can enter the area according to the identified color and the access right of the area, and if it is consistent, the position and access right information is output, and if it is not consistent, an alarm is given.

[0041] An information statistics module is configured to count the position and access right information of the person wearing the safety helmet in the image information, and record it in the background.

[0042] Further, the color identification module is embedded with an improved YOLOv5 target detection model and a ResNet network model, the YOLOv5 target detection model and the ResNet network model are connected, and the output port of the YOLOv5 target detection model is used as the input of the ResNet network model.

[0043] Further, in the YOLOv5 target detection model, a module for loading a spatial transformer is added in the feature extraction network; in the YOLOv5 target detection model, a strong attention SENet module is added behind the high-level convolution layer or the pooling layer of the Backbone network; in the YOLOv5 target detection model, a soft attention FPN module is added behind the high-level convolution layer or the pooling layer of the Neck network; and in the ResNet network model, an SENet module is added after the third Block of the ResNet-50.

[0044] A computer device includes a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that the computer program is the safety helmet specific area color recognition system based on deep learning.

[0045] The safety helmet specific area color recognition method and system based on deep learning has the following advantages compared with the prior art:

[0046] The technical scheme separates the collected image from the human target and the background, further inputs the improved YOLOv5 algorithm model, outputs it to the ResNet and classification network, adds the strong attention module SENet in the ResNet network to pay strong attention to the color of the specific area on the safety helmet, adds the priority module at the end of the whole network to further enhance the attention degree of the specific area, and improves the performance of the network; even in the condition of dark light, the color of the safety helmet main body and the specific area can be accurately recognized, which gives the basis for judging whether the staff is in the limited area.

[0047] The specific area in the technical scheme adopts a display screen, a rectangular electronic display screen with a length-width ratio of 2:1 is designed and installed at the center position in front of the brim of the safety helmet, and a battery and a wireless communication module are installed on the back of the display screen and are in communication connection with the display screen; the display screen is connected with the background server or other equipment, the color of the display screen can be controlled and changed through the background, and then the color is recognized to make a corresponding judgment on the right of way. By using the advantages of the display screen, the color displayed on the display screen in the specific area of the safety helmet can be maximally recognized under the condition of dark ambient light or unclear vision, so as to quickly and accurately judge the level of the right of way of the staff. In addition, the right of way of staff of each position is not unchangeable, and the color of the display screen can be temporarily adjusted through the background, so that the personnel scheduling and utilization are maximized, and the waste of resources can be reduced.

[0048] The technical scheme can accurately obtain the position and right of way of the staff by recognizing the color of the whole safety helmet and the specific area, and can ensure that unauthorized personnel cannot enter the specific area by distinguishing and controlling different areas. The color displayed on the display screen in the specific area of the safety helmet is used to determine whether the staff is working in the correct area and wearing the safety helmet according to the safety production specification. By dividing the right of way into different levels and distinguishing the level of the right of way according to the color of the specified area on the safety helmet, the right of way of personnel in different areas is further limited; different levels of the right of way of staff can be flexibly adjusted according to actual needs; the area control and safety management can be better realized, and only authorized personnel can enter the corresponding area, realizing the functions of "borderline alarm and personnel identification", and improving the safety of work. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is the specific flow chart of the whole method in embodiment 1.

[0050] Figure 2 is the adding of a space converter module in the YOLOv5 feature extraction network in embodiment 1.

[0051] Figure 3 is the improved YOLOv5 structure diagram in Example 1.

[0052] Figure 4 is the plan view of the access right of different areas of the chemical plant in the present application.

[0053] Figure 5 is the specific access right and staff position information diagram in Example 1.

[0054] Figure 6 is the main safety helmet and the display screen located at the center of the front brim of the safety helmet in Example 1.

[0055] Figure 7 is the overall system structure diagram of Example 2. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The described embodiments are only some of the embodiments of the present application, not all. Without departing from the design concept of the present application, various modifications and improvements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope of the present application. Example 1

[0057] As shown in Figure 1 , a specific area color recognition method for safety helmets based on deep learning, the specific recognition method includes the following steps:

[0058] Step 1: Obtain a video monitoring picture, and continuously obtain video frames of the obtained video.

[0059] Image extraction is performed on the required monitoring picture by an image acquisition device such as a monitoring camera, image acquisition is performed on scenes such as specific work areas or dangerous areas, and real-time image information is obtained.

[0060] Step 2: Collect image information, analyze the obtained video frames, and when a human target exists on the detected image information, cut out the area image at the position of the human target.

[0061] Real-time recognition of the monitoring picture detects whether there is a changing part, i.e., a moving object; if a moving object appears in the monitoring picture, the image of the part is extracted.

[0062] Step 3: Perform human target and background segmentation processing on the human region image, and uniformly transform the background pixels into preset background color pixels to obtain a human target image.

[0063] The human target and the background are segmented and processed to obtain a human target frame image, and the specific operation manner of distinguishing the human target and the background pixels is as follows: according to the collected image information, continuous frame images are obtained, so as to obtain the difference value of each pixel point in the current frame image and the previous frame image, the difference value is compared with a preset threshold value, when the difference value exceeds the preset threshold value range, the corresponding pixel point is regarded as a foreground image, otherwise, it is regarded as a background image, and the moving object image in the image information is obtained through the foreground image, that is, the human target.

[0064] In the present example, the human target and the background need to be segmented to reduce the interference of the background on the human target recognition, so as to maximize the accurate recognition of the human target and then recognize the safety hat and the position of the specific area-display screen of the safety hat. According to the collected image information, continuous frame images are obtained, and a "foreground and background segmentation" algorithm is used to obtain the difference value of each pixel point in the current frame image and the previous frame image, the difference value is compared with a preset threshold value, when the difference value exceeds the preset threshold value range, the corresponding pixel point is regarded as a foreground image, otherwise, it is regarded as a background image, and the moving object image in the image information is obtained through the foreground image, that is, the human target.

[0065] Step 4: input the human target area into a pre-established convolutional neural network model, and perform multi-target parallel detection on the image information according to the pre-set target classification; the specific operation manner is as follows:

[0066] The human target area image is input into a human detection model, and the human detection model is a target detection model based on YOLOv5, which can identify the overall safety hat and the position of the specific area on the safety hat worn by the human body.

[0067] In the present example, before inputting the YOLOv5 target detection model, the color space of the input image needs to be converted from RGB color space to HSV color space. For example, the color is decomposed into hue, saturation and lightness three components, wherein the hue component can well distinguish the types of colors, and the saturation and lightness components can reflect the intensity and brightness of the color, therefore, using the HSV color space can improve the accuracy of safety hat color recognition to a certain extent. After conversion, it can be input into the YOLOv5 target detection model to obtain the position of the overall safety hat and the specific area on the safety hat, as shown in Figure 4 , the overall safety hat and the position of the specific area-display screen in the center of the safety hat brim are recognized.

[0068] There may be a spatial deviation in the position of the extracted safety helmet and the specific area of the safety helmet, or there may be a certain deviation and error in the monitoring picture taken by the monitoring camera; therefore, in order to better extract the features of the specific area on the safety helmet, further spatial transformation is needed for the position of the currently extracted safety helmet and the specific area on the safety helmet.

[0069] The position of the safety helmet and the specific area on the safety helmet extracted by the target detection model of YOLOv5 can be basically determined, but for subsequent feature extraction of the specific area on the safety helmet, further spatial transformation is needed for the position of the currently extracted safety helmet and the specific area on the safety helmet, and a spatial transformer module needs to be added to the YOLOv5 feature extraction network to make corresponding spatial transformation of the spatial domain information of the position of the safety helmet and the specific area on the safety helmet in the image, so as to extract the key information. The specific spatial transformer module is as shown in Figure 2 .

[0070] Further improvement is made to the target detection model based on YOLOv5, and the specific improvement method is: applying attention mechanism in the target detection model of YOLOv5, using strong attention and soft attention to distinguish the attention of the position of the specific area on the safety helmet and the position of the safety helmet as a whole, specifically, adding a strong attention SENet module after the high-level convolution layer or pooling layer of the Backbone network of YOLOv5, and adding a soft attention FPN module after the high-level convolution layer or pooling layer of the Neck network of YOLOv5, so as to effectively improve the accuracy and robustness of the model. The improved YOLOv5 structure diagram is as shown in Figure 3 .

[0071] After obtaining the position, further identification and classification processing of the color of the safety helmet as a whole and the color of the specific area on the safety helmet are needed; the specific steps for obtaining the color information of the safety helmet as a whole and the specific area on the safety helmet include:

[0072] According to the improved YOLOv5-based target detection model, the accurate position of the safety helmet and the specific area on the safety helmet is obtained, and the output result is taken as the input of the ResNet network model to extract the features of the safety helmet and the specific area on the safety helmet, and finally the extracted features are input into the classification network for color classification.

[0073] The third Block of ResNet-50 in the ResNet network model is increased with an SENet module as a strong attention mechanism. The main purpose of the SENet module is to enhance the network's attention to specific features and improve network performance. Specifically, the position features of the safety helmet and the specific area on the safety helmet output by YOLOv5 are used, and these features are introduced into the SENet module to enhance the attention to the color of the safety helmet through the Squeeze-and-Excitation operation. The network can more focused on the features of the safety helmet color, thereby improving the accuracy and reliability of the identification. The purpose of selecting the SENet strong attention module here is to reduce the complexity of the entire network as much as possible.

[0074] In this example, the features extracted by the YOLOv5 model are abstracted and integrated at a higher level to obtain higher quality feature representations. An auxiliary network is added before the strong attention module of the ResNet model to predict the position information of the specific area of the safety helmet, and the prediction results are provided to the strong attention module. Specifically, an additional output can be added in YOLOv5, which will predict the position of the display screen of the safety helmet before the strong attention module. The strong attention module can use the specific area as a focus point to further identify the color on the safety helmet after positioning.

[0075] The color information of the safety helmet and the specific area of the safety helmet is extracted, and the specific steps include:

[0076] An auxiliary network is added before the strong attention module of the ResNet model to predict the position information of the specific area of the safety helmet, and the prediction results are provided to the strong attention module. Specifically, an additional output is added in YOLOv5, which will predict the position of the display screen of the safety helmet before the strong attention module; the strong attention module can use the specific area as a focus point to further identify the color on the safety helmet after positioning.

[0077] The color of the safety helmet and the specific area on the safety helmet is identified using a priority module to perform a weighted average algorithm to adjust the weight of the specific area, so that even if the color of the main body of the safety helmet cannot be accurately identified, the information can be judged by identifying the color of the specific area.

[0078] The color information of the safety helmet and the specific area on the safety helmet, the specific color requirements and the right to pass are as follows:

[0079] (I) Basic color

[0080] The color of the safety helmet in the chemical industry is generally regulated as follows:

[0081] 1) White: Project leader and workshop team management personnel;

[0082] 2) Blue: Special personnel and mechanical operation and maintenance personnel;

[0083] 3) Yellow: Construction team personnel;

[0084] 4) Red: Project full-time safety officer;

[0085] In order not to destroy the safety helmet color specified in the chemical industry, under the premise that the main color of the safety helmet does not change, a specific area is selected, a display screen is installed, the color of the display screen is changed, and the activity area of different personnel and the identity of the related personnel are distinguished.

[0086] (II) Area access color

[0087] 1) Highest access right: green; The specific area is green, which is matched with any of the above four theme colors, and the safety helmet with the highest access right can access the whole area;

[0088] 2) Medium access right: purple; The specific area is purple, which is matched with any of the above four theme colors, and the safety helmet with medium access right can access the medium and low level areas;

[0089] 3) Lowest access right: orange; The specific area is orange, which is matched with any of the above four theme colors, and the safety helmet with the lowest access right can only access the low level area. As shown in Figure 5 The structure of the safety helmet is shown in Figure 6 .

[0090] Step 5: According to the parallel detection result, the detection result at least includes the information of the overall color of the safety helmet and the position and color of the specific area on the safety helmet.

[0091] Step 6: According to the parallel detection result, the situation of personnel wearing safety helmet in the collected image information is judged, the human body target image is input into the YOLOv5 target detection model, the safety helmet position is recognized, if the safety helmet position is not recognized, it is judged that the worker does not wear safety helmet, and an alarm information is sent.

[0092] Specifically, the position of the safety helmet and the specific area-display screen on the safety helmet is obtained, if the safety helmet position is not recognized, it is judged that the worker does not wear safety helmet, and an alarm information is sent; If the personnel who correctly wear the safety helmet are detected, they can be input into the subsequent ResNet and classification network, the color of the safety helmet and the specific area-display screen on the safety helmet is recognized, and according to the color of the display screen, whether the access right of the worker conforms to the color of the specified area is judged, if it conforms, it can enter, if it does not conform, an alarm information is sent.

[0093] As shown in Figure 4 The chemical plant temporarily defines the dangerous area and the maintenance area as the area that the personnel with high access right can enter, the specific area as the area that the personnel with medium access right can enter, and the working area as the area that the personnel with low access right can enter. The bottom color of the safety helmet of the personnel in the specific area is blue, and the display screen of the safety helmet is purple. Through the combination of information, it can be judged that the personnel is a mechanical operation and maintenance personnel with medium access right, and the personnel usually works in the specific area. If an abnormality occurs in the working area at this time and the personnel needs to perform on-site maintenance, since the access right of the personnel is medium, the personnel can enter the medium or low level area, and thus the personnel can directly enter the working area to perform on-site maintenance. If an abnormality occurs in the dangerous area at this time, and the personnel needs to be urgently mobilized to perform on-site investigation, the personnel needs to be changed in access right by the personnel in the background or the computer program, so that the color of the display screen of the safety helmet is changed to green, so as to achieve the highest level of access right. In this way, the personnel entering the dangerous area can be reduced, and the working area of the personnel can be standardized, so as to realize the standard and safety of industrial production.

[0094] Example 2

[0095] As shown in Figure 7 A safety helmet specific area color recognition system based on deep learning can be applied to the safety helmet specific area color recognition method based on deep learning. The recognition system comprises:

[0096] An image processing module acquires image information and separates human targets and backgrounds from moving object images in the image information.

[0097] A position recognition module is used to perform multi-target parallel detection on YOLOv5, and obtain the position of the safety helmet and the display screen. The detection result at least includes the overall color of the safety helmet and the position information of the specific area on the safety helmet.

[0098] A position determination module is used to determine whether the personnel wears a safety helmet according to the parallel detection result of the personnel wearing a safety helmet in the collected image information. If the position of the safety helmet is not recognized, it is determined that the personnel does not wear a safety helmet, and an alarm is given.

[0099] A color recognition module performs color recognition on ResNet and a classification network according to the input, and performs classification.

[0100] A color determination module determines whether the access right level of the personnel is allowed to enter the area according to the recognized color and the access right of the area. If it is allowed, the position and access right information are output. If it is not allowed, an alarm is given.

[0101] An information statistics module is configured to count the information of the position and the right of passage of the person wearing the safety helmet in the image information and record the information in the background.

[0102] The color recognition module is embedded with an improved YOLOv5 target detection model and a ResNet network model, the YOLOv5 target detection model and the ResNet network model are connected, and the output port of the YOLOv5 target detection model is used as the input of the ResNet network model.

[0103] In the YOLOv5 target detection model, a module for loading a spatial transformer is added in the feature extraction network; in the YOLOv5 target detection model, a strong attention SENet module is added behind a high-level convolution layer or a pooling layer of a Backbone network; in the YOLOv5 target detection model, a soft attention FPN module is added behind a high-level convolution layer or a pooling layer of a Neck network; and in the ResNet network model, an SENet module is added behind the third Block of a ResNet-50. Embodiment 3

[0104] A computer device comprises a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, the computer program being the recognition system for the specific area color of the safety helmet based on deep learning described in Embodiment 2; and the processor executes the computer program to implement the recognition method for the specific area color of the safety helmet based on deep learning described in Embodiment 1.

Claims

1. A method for identifying the color of a specific area of a safety helmet based on deep learning, characterized in that, The specific identification method comprises the following steps: Step 1: Obtain a video monitoring picture, and continuously acquire video frames of the obtained video; Step 2: Collect image information, analyze the obtained video frames, and when a human body target is detected on the image information to be measured, intercept a region image at a position of the human body target; Step 3: Perform human body target and background segmentation processing on the human body region image, and uniformly transform background pixels into preset background color pixels to obtain a human body target image; Step 4: Input the human body target region into a pre-established convolutional neural network model, and perform multi-target parallel detection on the image information according to a pre-set target classification; the specific operation mode is as follows: input the human body target region image into a human body detection model, the human body detection model is a target detection model based on YOLOv5, and the overall safety helmet worn by the human body and the position of the specific region on the safety helmet are recognized; The position of the overall safety helmet and the specific region on the safety helmet extracted by the target detection model of YOLOv5 can be determined, but for subsequent feature extraction of the specific region on the safety helmet, further spatial transformation is required for the position of the overall safety helmet and the specific region on the safety helmet currently extracted, and a spatial transformer module needs to be added in the YOLOv5 feature extraction network, so as to perform corresponding spatial transformation on the position of the safety helmet and the spatial domain information of the specific region on the safety helmet in the image, so as to extract the key information; Step 5: According to the parallel detection result, a detection result is obtained, which at least includes information of the overall safety helmet color and the specific region position and color on the safety helmet; the specific steps for obtaining the information of the overall safety helmet color and the specific region color on the safety helmet include: After obtaining the human body target region, further identification and classification processing of the overall safety helmet color and the specific region color on the safety helmet are required; According to the improved YOLOv5-based target detection model, the accurate position of the safety helmet and the specific region on the safety helmet is obtained, and the output result is used as the input of the ResNet network model to extract the features of the safety helmet and the specific region on the safety helmet, and finally the extracted features are input into a classification network for color classification; An SENet module is added after the third Block of ResNet-50 in the ResNet network model as a strong attention mechanism, and the main purpose of the SENet module is to enhance the attention of the network to specific features and improve the network performance; specifically, the position features of the safety helmet and the specific region on the safety helmet output by YOLOv5 are used, and these features are introduced into the SENet module to enhance the attention to the safety helmet color through Squeeze-and-Excitation operation.

2. The method for identifying the color of a specific area of a safety helmet based on deep learning according to claim 1, characterized in that: Further improvements are made to the YOLOv5-based target detection model, and the specific improvement method is: applying attention mechanism in the YOLOv5 target detection model, using strong attention and soft attention to distinguish the attention of the specific area on the safety helmet and the overall position of the safety helmet, specifically, adding a strong attention SENet module after the high-level convolution layer or pooling layer of the Backbone network of YOLOv5, and adding a soft attention FPN module after the high-level convolution layer or pooling layer of the Neck network of YOLOv5, so as to effectively improve the accuracy and robustness of the model.

3. The method of claim 1, wherein the method is based on deep learning. The color information of the safety helmet and the specific area of the safety helmet is extracted, and the specific steps include: An auxiliary network is added before the strong attention module of the ResNet model to predict the position information of the specific area of the safety helmet, and the prediction result is provided to the strong attention module, specifically, an additional output is added in YOLOv5, which will predict the display screen position of the safety helmet before the strong attention module; in this way, the strong attention module will further identify the color on the base color of the safety helmet after positioning the specific area as the attention point; The color of the safety helmet and the specific area on the safety helmet is identified using a priority module to perform a weighted average algorithm to adjust the weight of the specific area, so that even if the color of the safety helmet body cannot be accurately identified, the information can be judged by identifying the color of the specific area.

4. The method according to any one of claims 1-3, wherein the method is based on deep learning to identify the color of a specific area of the safety helmet. The information of the overall color of the safety helmet and the specific area on the safety helmet, the specific color requirements and the passing right situation are as follows: (1) Basic color The color of the safety helmet in the chemical industry is generally as follows: 1) White: project leaders and workshop team management personnel; 2) Blue: special personnel and mechanical operation and maintenance personnel; 3) Yellow: construction team personnel; 4) Red: project full-time safety officers; In order to not destroy the safety helmet color specified in the chemical industry, the specific area is selected on the premise that the main color of the safety helmet does not change, the display screen is installed, the color of the display screen is changed, and the activity area of different personnel and the identity of the related personnel are distinguished; (2) Regional passing color 1) Highest passing right: green; the specific area is green, which is matched with any of the above four theme colors, and the safety helmet has the highest passing right and can pass through the whole area; 2) Medium passing right: purple; the specific area is purple, which is matched with any of the above four theme colors, and the safety helmet has a medium passing right and can pass through the medium and low level area in the safety level requirement; 3) Lowest passing right: orange; the specific area is orange, which is matched with any of the above four theme colors, and the safety helmet has the lowest passing right and can only pass through the low level area in the safety level requirement.

5. The method of claim 1, wherein the method is based on deep learning. The specific operation mode of the human body target and background segmentation processing in step 3 is as follows: according to the collected image information, continuous frame images are obtained, so that the difference values of each pixel point in the current frame image and the previous frame image are obtained, the difference values are compared with a preset threshold value, when the difference values exceed the preset threshold value range, the corresponding pixel points are regarded as foreground images, otherwise, they are regarded as background images, and the moving object image in the image information is obtained through the foreground image, that is, the human body target.

6. A system for identifying the color of a specific area of a safety helmet based on deep learning, applicable to the method for identifying the color of a specific area of a safety helmet based on deep learning according to any one of claims 1-5, characterized in that: The recognition system comprises: An image processing module acquires the moving object image in the image information, and separates the human body target and the background therefrom; A position recognition module is configured to perform multi-target parallel detection on the input to the YOLOv5, and to obtain the position of the safety helmet and the display screen, wherein the detection result at least comprises the overall color of the safety helmet and the position information of the specific area on the safety helmet; A position determination module is configured to determine whether the safety helmet is worn by the personnel in the collected image information according to the parallel detection result, and to alarm if the safety helmet position is not recognized; A color recognition module is configured to perform color recognition on the input to the ResNet and the classification network, and to perform classification; A color determination module is configured to determine whether the personnel holding the access right can enter the area according to the recognized color and the access right of the area, and to output the position and access right information if the determination is positive, and to alarm if the determination is negative; An information statistics module is configured to count the position and access right information of the personnel wearing the safety helmet in the image information, and to record the information in the background.

7. The system according to claim 6, wherein the system is based on deep learning to identify the color of the specific area of the safety helmet. The color recognition module is embedded with the improved target detection model of YOLOv5 and the ResNet network model, the target detection model of YOLOv5 and the ResNet network model are connected, and the output port of the target detection model of YOLOv5 is used as the input of the ResNet network model.

8. The system according to claim 7, wherein the system is based on deep learning to identify the color of the specific area of the safety helmet. In the target detection model of YOLOv5, a module for loading a spatial transformer is added in the feature extraction network; in the target detection model of YOLOv5, a strong attention SENet module is added behind the high-level convolution layer or the pooling layer of the Backbone network; in the target detection model of YOLOv5, a soft attention FPN module is added behind the high-level convolution layer or the pooling layer of the Neck network; and in the ResNet network model, an SENet module is added behind the third Block of the ResNet-50.

9. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that: The computer program is the recognition system of the specific area color of the safety helmet based on deep learning according to any one of claims 6-8; and the processor executes the computer program to realize the recognition method of the specific area color of the safety helmet based on deep learning according to any one of claims 1-5.

Images