Goggle miswear detection method

Abstract

The application relates to the technical field of face detection, in particular to a goggle incorrect wearing detection method, which comprises the following steps: face detection; goggle state detection; the specific steps of face detection comprise the following steps: acquiring a face sample image, preparing a face sample set; designing a face detection main network to detect the face sample set; a double-step detection algorithm is designed aiming at factors such as multiple goggle types, multiple positions and particularity, first, a face is correctly detected, and then, a goggle state is detected; the improved SSD is used to find out a face in an image, then a classification network is used to identify whether the face is correctly worn, and the area of the original image where the goggle is correctly worn can be detected; the application provides a very strong detection method which can accurately detect whether the goggle is correctly worn on the face.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of face detection, and in particular to a method for detecting incorrect wearing of goggles. BACKGROUND

[0002] Goggles are a protection for workers in harmful and toxic posts, and safety work specifications require that workers must wear goggles correctly. The principles of correct wearing of goggles are: 1. The required type of goggles must be worn according to requirements; 2. Goggles must be correctly worn on the eyes; and 3. Goggles cannot be damaged.

[0003] Since there is generally video monitoring on the post, it is feasible to detect incorrect wearing of goggles based on video images. The essence of detecting incorrect wearing of goggles based on images is an image target detection problem. The most common target detection method based on deep learning at present is a single-stage SSD, a two-stage Yolo series, and FasterRCNN. The three methods have been widely applied to other target detection, but the detection of incorrect wearing of goggles has certain particularity. The detection object can only be goggles on the face, and goggles in other places cannot be correctly worn even if they are of the correct type and are not damaged. Therefore, a more rigorous detection logic relationship needs to be designed to complete this work. At present, there is no detection method for whether goggles are correctly worn. SUMMARY

[0004] The purpose of the present application is to provide a method for detecting incorrect wearing of goggles, which can detect whether goggles are correctly worn on a face.

[0005] To achieve the above purpose, the present application provides a method for detecting incorrect wearing of goggles, comprising:

[0006] detecting a face;

[0007] detecting a state of goggles;

[0008] The specific steps of detecting a face include:

[0009] obtaining a face sample image to make a face sample set;

[0010] designing a face detection backbone network to detect the face sample set.

[0011] The specific steps of detecting a state of goggles include:

[0012] According to the design and model training of samples, face detection is performed in a sample library to obtain detection results of images of different sizes, and a candidate frame correctly detecting a face is selected;

[0013] Determine the size of the classifier input image, and balance the size of the selected candidate box to make the size of the candidate box consistent with the size of the classifier input image before entering the classifier training;

[0014] Resnet50 network is built for deep learning, and the network model is saved when the accuracy reaches 99% after at least 20 training batches.

[0015] The face sample image includes a face image collected from the network and a face image shot on site.

[0016] The face sample image is specifically: the image of the face with goggles, no goggles, damaged goggles, sunglasses, helmet, mask, and face with occlusion.

[0017] The face sample set includes positive samples and negative samples; the positive sample label is a face rectangle, whether the face has occlusion or not, whether the face has goggles or not; the negative sample is all images that are not face backgrounds.

[0018] The ratio of the positive sample to the negative sample is not less than 1:3.

[0019] The goggle incorrect wearing detection method of the application is designed for the factors such as multiple types of goggles, special positions, etc., and a two-step detection algorithm is designed, which first correctly detects the face and then detects the goggle state. The application first uses the improved SSD to find the face in the image, and then uses the classification network to identify whether the face is correctly worn or not. The region of the original image where the goggle is correctly worn can be detected, such as the model trained to detect the ABCD four types of goggle. If only B type goggle is needed to be detected in a certain occasion, ACD correct wearing and incorrect wearing are all regarded as violations. The application provides a very strong detection method. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0021] Figure 1 is a flowchart of the goggle incorrect wearing detection method of the application.

[0022] Figure 2 is a flowchart of detecting the face of the application.

[0023] Figure 3A flowchart of detecting the state of the protective glasses according to the present application.

[0024] Figure 4 A structural schematic diagram of the face detection backbone network according to the present application. DETAILED DESCRIPTION

[0025] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0026] Referring to Figures 1-4 , the present application provides a protective glasses incorrect wearing detection method, comprising:

[0027] S1, detecting a face;

[0028] The specific steps include:

[0029] S11, obtaining a face sample image, and making a face sample set;

[0030] According to the face images collected on the network and the face images shot on the spot, the mixed quantity is about 10,000; the size of the face is from 300*300 to 2*2, and the number of each size is basically equivalent; the face has a certain number of situations such as wearing protective glasses, not wearing protective glasses, damaged protective glasses, wearing sunglasses, wearing a helmet, wearing a mask, and face occlusion; positive sample: the positive sample label is a face rectangle, whether the face has occlusion or not, whether the face has protective glasses or not; negative sample: all the background that is not the face. Since there are too many negative samples, the positive and negative sample ratio is set to not less than 1:3.

[0031] S12, designing a face detection backbone network to detect the face sample set;

[0032] The face detection backbone network is a two-classification network, the type is relatively easy to judge, and the key is to accurately determine the position parameter. The designed backbone network based on SSD is as shown in Figure 4 According to the standard SSD network, a new network structure based on recursive structure is designed, the new network structure considers the influence factor of the neighborhood, and does not use the overall pyramid structure, but fuses two adjacent feature layers in the down-sampling for a certain scale feature layer, so as to complete the feature fusion of the current feature layer, and enhance the detection accuracy.

[0033] The specific network is described as follows:

[0034] a) the input is a color image of 300*300*3;

[0035] b) The main layer adopts the model of VGG16 of standard SSD, and L1 to L6 main feature layers are obtained respectively;

[0036] c) For L2 layer, it is copied into 2 paths; one path is sampled into 38*38*256 m11, and the other path is unchanged; for L3 layer, it is copied into 3 paths, one path is sampled into 19*19*1024 feature layer m21, the other path is sampled into 38*38*256 m12, and the other path is unchanged L3; for L4 layer, it is copied into 3 paths, one path is sampled into 10*10*512 feature layer m31, the other path is sampled into 19*19*1024 m22; for L5 layer, it is copied into 3 paths, one path is sampled into 10*10*512 feature layer m32, the other path is sampled into 3*3*256 m41, and the other path is unchanged; for L6 layer, it is copied into 2 paths, one path is sampled into 5*5*256 feature layer m42, and the other path is unchanged.

[0037] d) For the detector and classifier 1, the input is L1+0.5(m11+m12); for the detector and classifier 2, the input is L2+0.5(m21+m22); for the detector and classifier 3, the input is L3+0.5(m31+m32); for the detector and classifier 4, the input is L4+0.5(m41+m42); for the detector and classifier 5, the input is L5, and for the detector and classifier 6, the input is L6; according to the design, the feature layer of small scale is not changed, and the recursive structure is increased for the feature layer of large scale.

[0038] e) A fast non-maximum suppression module is connected behind the feature detection and classifier to complete the design of the network.

[0039] S2 detects the state of the goggles;

[0040] The specific steps include:

[0041] S21 detects the face according to the sample design and model training, obtains the detection results of images of different sizes from the sample library, and selects the candidate frame of the correctly detected face;

[0042] According to the sample design and model training, the face is detected from the sample library, and the detection results of images of different sizes are obtained. Only the candidate frame of the correctly detected face is selected, and the picture of the candidate frame can be all face candidate frames correctly and incorrectly wearing goggles. The coordinates of each candidate frame in the original picture are (x, y, w, h), x and y are the left top points; w and h are the width and height of the frame.

[0043] S22 determines the size of the classifier input image, and balances the size of the selected candidate frame, so that the size of the candidate frame is consistent with the size of the classifier input image, and then enters the classifier training;

[0044] The size w, h of the face candidate frame obtained through step S22 is balanced, that is, the average values w' and h' of w and h are calculated, and the size w' and h' of the classifier input image is determined.

[0045] S23 builds a Resnet50 network for deep learning, and the training batch is at least 20 times, and when the accuracy rate reaches 99%, the network model is saved.

[0046] A Resnet50 network is built for deep learning, and the training batch is at least 20 times, and the key parameters of the model can be set regularly. When the accuracy rate reaches 99%, the network model is saved. The multi-classification problem can easily reach this accuracy.

[0047] The goggle correct wearing detection method of the present application is designed for the factors such as multiple types of goggles, special positions and the like, and a two-step detection algorithm is designed, which first correctly detects a face and then detects a goggle state. The present application first finds a face in an image by using an improved SSD, and then identifies whether the face is correctly worn by using a classification network, so that the area of the original image where the goggle is correctly worn can be detected. If the model is trained to detect the cases of four types of ABCD goggle, if only the B type goggle is needed to be detected in a certain occasion, the correct wearing and incorrect wearing of ACD goggle are all regarded as violations, and the present application provides a very strong detection method.

[0048] The above only discloses one preferred embodiment of the present application, and of course cannot limit the scope of the rights of the present application, and those skilled in the art can understand that all or part of the above-mentioned embodiments can be implemented, and equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.

Claims

1. A method of eyewear mis-wearing detection, the method comprising: The method comprises the following steps: detecting a face; detecting a goggle state; The specific steps of detecting the face include: obtaining a face sample image and making a face sample set; designing a face detection backbone network to detect the face sample set; wherein the face detection backbone network is a binary classification network, and is a backbone network based on SSD. The specific network description of the face detection backbone network is as follows: the input is a 300*300*3 color image; the backbone layer adopts a model of VGG16 of a standard SSD, and obtains L1 to L6 main feature layers; for the L2 layer, 2 paths are copied; one path is sampled to obtain m11 of 38*38*256, and the other path remains unchanged; for the L3 layer, 3 paths are copied, one path is sampled to obtain a feature layer m21 of 19*19*1024, the other path is sampled to obtain m12 of 38*38*256, and the other path remains unchanged L3; for the L4 layer, 3 paths are copied, one path is sampled to obtain a feature layer m31 of 10*10*512, the other path is sampled to obtain m22 of 19*19*1024; for the L5 layer, 3 paths are copied, one path is sampled to obtain a feature layer m32 of 10*10*512, the other path is sampled to obtain m41 of 3*3*256, and the other path remains unchanged; for the L6 layer, 2 paths are copied, one path is sampled to obtain a feature layer m42 of 5*5*256, and the other path remains unchanged; for a detector and a classifier 1, the input is L1+0.5(m11+m12); for a detector and a classifier 2, the input is L2+0.5(m21+m22); for a detector and a classifier 3, the input is L3+0.5(m31+m32); for a detector and a classifier 4, the input is L4+0.5(m41+m42); for a detector and a classifier 5, the input is L5; for a detector and a classifier 6, the input is L6; according to the design, the feature layer of a small scale is not changed, and the recursive structure is added to the feature layer of a large scale; a fast non-maximum suppression module is connected behind the feature detector and the classifier, and the design of the network is completed.

2. The goggle incorrect wearing detection method according to claim 1, wherein the specific steps of detecting the goggle state include: According to the sample design and model training, face detection is performed on the sample library to obtain the detection results of images of different sizes, and the candidate boxes correctly detecting the face are selected; The size of the classifier input image is determined, the selected candidate box size is balanced, the candidate box size is consistent with the size of the classifier input image, and then the balanced candidate box enters the classifier training; A Resnet50 network is built for deep learning, and the network model is saved when the accuracy reaches 99% after at least 20 batches of training.

3. The goggle incorrect wearing detection method according to claim 1, wherein the face sample image includes face images collected from the network and face images shot on site.

4. The goggle incorrect wearing detection method according to claim 1, wherein ​ ​ The face sample image is specifically: an image of a face with goggles, no goggles, damaged goggles, sunglasses, wearing a helmet, wearing a mask, and a face with an occlusion. 5.The method of claim 1, wherein, The face sample set includes positive samples and negative samples; the positive sample label is a face rectangle, regardless of whether the face has an occlusion or has goggles; and the negative sample is a background that is not a face in all images. 6.The method of claim 5, wherein, The ratio of the positive samples to the negative samples is not less than 1:3.

Images