Natural scene high-precision face recognition method applied to golf course

Abstract

The invention relates to the technical field of face recognition, in particular to a natural scene high-precision face recognition method applied to a golf course, and the method comprises the following steps: collecting a golf course data training model under different illumination conditions; acquiring a high dynamic range image by using an HDR camera; preprocessing the image; face recognition is started; the face recognition comprises the following steps: data input; detecting a human face; aligning the human faces; extracting face features; and face matching. Compared with the prior art, the natural scene high-precision face recognition method applied to the golf course can better adapt to strong light and weak light environments; image preprocessing, image filtering and denoising, color correction and other operations can maintain the quality of the image; by means of multi-scale feature extraction, face information of different sizes can be captured, and therefore it can be ensured that the face can be accurately detected regardless of the size of the face in the image.

Description

[Technical field]

[0001] The present invention relates to the field of face recognition technology, and in particular to a high-precision face recognition method for natural scenes applied to golf courses. [Background technology]

[0002] With the increasing demand for intelligent golf courses, facial recognition technology is widely used in scenarios such as golf course management, member identification, and security monitoring.

[0003] The application of existing face recognition technology in natural scenes is limited by factors such as ambient light and weather conditions, and cannot operate stably in strong light, weak light and bad weather. In addition, existing face recognition technology requires the object to be recognized to be in a specified position or maintain a specific posture, which cannot meet the use requirements in the open environment of a golf course. [Summary of the invention]

[0004] In order to overcome the above problems, the present invention proposes a high-precision face recognition method for natural scenes applied to golf courses, which can effectively solve the above problems. The method is specifically used for golf courses and can maintain high accuracy in strong light and weak light environments, and the recognizer may not be aware of the recognition process.

[0005] The present invention provides a technical solution to solve the above technical problems: providing a high-precision face recognition method for natural scenes applied to golf courses, comprising the following steps:

[0006] Step S1, collecting golf course data under different lighting conditions to train a model;

[0007] Step S2, using an HDR camera to acquire a high dynamic range image;

[0008] Step S3, image preprocessing;

[0009] Step S4, start face recognition;

[0010] The step S4 includes the following steps:

[0011] Step S41, data input;

[0012] Step S42, face detection;

[0013] Step S43, face alignment;

[0014] Step S44, facial feature extraction;

[0015] Step S45: face matching.

[0016] Preferably, in step S44, a face recognition model is used to extract feature vectors of the aligned face images.

[0017] Preferably, in step S44, extracting a feature vector of the aligned face image using a face recognition model comprises the following steps:

[0018] Step S441, constructing an image pyramid, starting from the original image, gradually reducing the image resolution, and constructing a multi-level image pyramid, wherein the resolution of each layer of the image is gradually reduced by half each time;

[0019] Step S442, using the convolutional neural network vgg to extract features on each layer of the image;

[0020] Step S443: Features of different scales are merged using a feature concatenation method to form a multi-scale feature representation.

[0021] Step S444: sending the multi-scale features to the detection head of the face recognition model to detect the face position.

[0022] Preferably, in step S44, the arcFace model is used to extract facial features, and a floating-point array with a dimension of 512 is extracted as a unique feature vector of the face.

[0023] Preferably, the step S1 includes the following steps:

[0024] Step S11, more than ten groups of cameras are arranged on the golf course to collect images at all times of the day, including morning, noon, and dusk;

[0025] Step S12, collecting images in different weather conditions;

[0026] Step S13, filtering out images containing people participating in golf from the collected images;

[0027] Step S14: Use the labeling tool labelme to label the face area to ensure that the labeled faces cover different lighting scenes.

[0028] Preferably, in step S3, the collected high dynamic range image is preprocessed, including image alignment, noise removal, and color correction.

[0029] Preferably, the HDR camera adopts a Sony IMX678 image sensor to acquire a high dynamic range image through the Sony IMX678 image sensor.

[0030] Preferably, in step S41, the data input is a high dynamic range image captured by the IMX678 image sensor. After image preprocessing, the yoloFace model is used to detect faces in the image and output a face rectangular frame; the face image is cut out through the face rectangular frame as input for face alignment.

[0031] Preferably, in step S43, the key point detection model detects key points of the face and aligns the face image to a standard posture through affine transformation.

[0032] Preferably, in step S45, the extracted feature vector is compared with known facial features in the database, the cosine similarity method is used to calculate the similarity between the input face and each face in the database, and the result with the highest similarity is selected as the recognition result.

[0033] Compared with the prior art, the high-precision face recognition method for natural scenes applied to golf courses of the present invention collects golf course data under different lighting conditions to train the model. The model can be more stable in different lighting scenes and better adapt to strong light and weak light environments; an HDR camera is used to obtain high dynamic range images. HDR will continuously shoot images, respectively taking images in underexposure, normal exposure and overexposure scenes, and then combining these scene images together and highlighting the best part of each photo to generate a clear image. HDR makes the model more adaptable to strong light and weak light environments; image preprocessing, image filtering, denoising, color correction and other operations can maintain the quality of the image, and the quality of the image affects the stability of the algorithm model; multi-scale feature extraction can capture face information of different sizes. This ensures that no matter how big or small the face is in the image, it can be accurately detected.

Brief Description of the Drawings

[0034] Figure 1 This is an overall flow chart of the high-precision face recognition method for natural scenes applied to golf courses of the present invention;

[0035] Figure 2 A face recognition flow chart of the high-precision face recognition method for natural scenes applied to a golf course according to the present invention;

[0036] Figure 3 This is a feature map for extracting facial features of a natural scene high-precision face recognition method applied to a golf course by the present invention. [Specific implementation method]

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and implementation examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0038] It should be noted that in the embodiments of the present invention, all directional indications (such as up, down, left, right, front, back, etc.) are limited to relative positions on the specified view, rather than absolute positions.

[0039] In addition, in the present invention, descriptions such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0040] See also Figures 1 to 3 The high-precision face recognition method for natural scenes applied to golf courses of the present invention is specifically applied to golf courses and includes the following steps:

[0041] Step S1, collecting golf course data under different lighting conditions to train a model.

[0042] The step S1 includes the following steps:

[0043] Step S11, more than ten groups of cameras are arranged on the golf course to collect images at all times of the day, including morning, noon, and dusk;

[0044] Step S12, collecting images in different weather conditions, such as cloudy days and sunny days;

[0045] Step S13, filtering out images containing people participating in golf from the collected images;

[0046] Step S14: Use the labeling tool labelme to label the face area to ensure that the labeled faces cover different lighting scenes.

[0047] Step S2, using an HDR camera to acquire a high dynamic range image (HDR image).

[0048] In step S2, the HDR camera can capture a wide range of light intensities, thereby adapting to strong light and weak light environments.

[0049] The HDR camera adopts the Sony IMX678 image sensor to obtain high dynamic range images. Regardless of strong light, weak light, front light or back light, clear images can be obtained during the day.

[0050] Step S3: image preprocessing.

[0051] In step S3, the collected high dynamic range image (HDR image) is preprocessed, including image alignment, noise removal, color correction, etc., to ensure image quality and consistency.

[0052] In step S3, image denoising is completed by performing Gaussian filtering using opencv. Opencv is a powerful and feature-rich image operation library, and opencv implements the GaussianBlur interface to provide Gaussian filtering operations.

[0053] In step S3, CLAHE (contrast limited adaptive histogram equalization) is used to perform image color correction. The CLAHE object of opencv provides an interface of the CLAHE method, and the image color correction is completed using the interface.

[0054] Step S4, start face recognition.

[0055] The step S4 includes the following steps:

[0056] Step S41, data input;

[0057] Step S42, face detection;

[0058] Step S43, face alignment;

[0059] Step S44, facial feature extraction;

[0060] Step S45, face matching;

[0061] Step S46, result processing.

[0062] In the step S41, the data input is a high dynamic range image captured by the IMX678 image sensor. After image preprocessing, the yoloFace model is used to detect the face in the image, and a face rectangular frame (x1, y1, x2, y2) is output; the face image is cut out through the face rectangular frame as the input for face alignment.

[0063] In step S42, a deep learning model is used to detect the position of the face in the image.

[0064] In step S43, the key point detection model detects key points of the face and aligns the face image to the standard posture through affine transformation. 68 key points on the face including eyes, nose, mouth, etc. are detected using dlib, and then the face in the image is adjusted to the standard posture through the transformation matrix.

[0065] In step S44, a face recognition model is used to extract a feature vector of the aligned face image. The feature vector is a high-dimensional floating point array representing the unique features of the face.

[0066] In step S44, the arcFace model is used to extract facial features, and a floating point array with a dimension of 512 is extracted as a unique feature vector of the face.

[0067] In step S44, multi-scale feature extraction is used to extract image features at different scales.

[0068] In step S44, extracting feature vectors of the aligned face images using a face recognition model includes the following steps:

[0069] Step S441, constructing an image pyramid: starting from the original image, gradually reducing the image resolution to construct a multi-level image pyramid. The resolution of each layer of the image is gradually reduced by half each time.

[0070] Step S442, feature extraction network: use convolutional neural network vgg to extract features on each layer of image.

[0071] Step S443, feature fusion: features of different scales are fused using a feature concatenation method to form a multi-scale feature representation.

[0072] Step S444: sending the multi-scale features to the detection head of the face recognition model to detect the face position.

[0073] See also Figure 3 , the bottom of the figure is the input image. From bottom to top, there are five convolutional layers (conv1 to conv5), each followed by a feature layer (C1 to C5). These convolutional layers are used to extract features of the image, and as the network goes deeper, the size of the feature map gradually decreases, while the number of channels gradually increases. The stride of each convolutional layer gradually increases (from 2 to 32), which leads to a gradual decrease in the size of the feature map. These are the outputs of the convolutional layers, which will be used to build the feature pyramid. The bottom-up path, that is, the direct connection from the input image to the high-level features. The top-down path, that is, the connection from the high-level features to the low-level features, is usually implemented by upsampling (such as 0.5x represents upsampling). The lateral connection, that is, the connection between different layers, is usually implemented by 1x1 convolution, which is used to fuse features of different scales. In the figure, M2 to M5 represent feature fusion layers, which receive features from the lower layers through lateral connections and pass them to higher layers through the top-down path. Output feature maps (P2-P5): These are the final output feature maps, which have different scales and can be used to detect objects of different sizes. 3x3 convolution: After each M layer, there is a 3x3 convolution layer to further process the fused features.

[0074] In step S45, the extracted feature vector is compared with known facial features in the database to identify the identity of the person.

[0075] In step S45, the extracted feature vector is compared with known face features in the database, and the similarity between the input face and each face in the database is calculated using the cosine similarity method, and the result with the highest similarity is selected as the recognition result.

[0076] Compared with the prior art, the high-precision face recognition method for natural scenes applied to golf courses of the present invention collects golf course data under different lighting conditions to train the model. The model can be more stable in different lighting scenes and better adapt to strong light and weak light environments; an HDR camera is used to obtain high dynamic range images. HDR will continuously shoot images, respectively taking images in underexposure, normal exposure and overexposure scenes, and then combining these scene images together and highlighting the best part of each photo to generate a clear image. HDR makes the model more adaptable to strong light and weak light environments; image preprocessing, image filtering, denoising, color correction and other operations can maintain the quality of the image, and the quality of the image affects the stability of the algorithm model; multi-scale feature extraction can capture face information of different sizes. This ensures that no matter how big or small the face is in the image, it can be accurately detected.

[0077] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any modifications, equivalent substitutions and improvements made within the concept of the present invention should be included in the patent protection scope of the present invention.

Claims

1. A high-precision face recognition method for natural scenes applied to golf courses, characterized in that: The steps include: Step S1, collecting golf course data under different lighting conditions to train a model; Step S2, using an HDR camera to acquire a high dynamic range image; Step S3, image preprocessing; Step S4, start face recognition; The step S4 includes the following steps: Step S41, data input; Step S42, face detection; Step S43, face alignment; Step S44, facial feature extraction; Step S45: face matching.

2. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: In step S44, a face recognition model is used to extract feature vectors of the aligned face images.

3. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 2, characterized in that: In step S44, extracting feature vectors of the aligned face images using a face recognition model includes the following steps: Step S441, constructing an image pyramid, starting from the original image, gradually reducing the image resolution, and constructing a multi-level image pyramid, wherein the resolution of each layer of the image is gradually reduced by half each time; Step S442, using the convolutional neural network vgg to extract features on each layer of the image; Step S443: Features of different scales are merged using a feature concatenation method to form a multi-scale feature representation. Step S444: sending the multi-scale features to the detection head of the face recognition model to detect the face position.

4. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: In step S44, the arcFace model is used to extract facial features, and a floating point array with a dimension of 512 is extracted as a unique feature vector of the face.

5. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: The step S1 includes the following steps: Step S11, more than ten groups of cameras are arranged on the golf course to collect images at all times of the day, including morning, noon, and dusk; Step S12, collecting images in different weather conditions; Step S13, filtering out images containing people participating in golf from the collected images; Step S14: Use the labeling tool labelme to label the face area to ensure that the labeled faces cover different lighting scenes.

6. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: In step S3, the collected high dynamic range image is preprocessed, including image alignment, noise removal, and color correction.

7. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: The HDR camera uses a Sony IMX678 image sensor to acquire high dynamic range images.

8. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 7, characterized in that: In the step S41, the data input is a high dynamic range image captured by the IMX678 image sensor. After the image is preprocessed, the yoloFace model is used to detect the face in the image and output the face rectangular frame; the face image is cut out through the face rectangular frame as the input for face alignment.

9. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: In step S43, the key point detection model detects key points of the face and aligns the face image to a standard posture through affine transformation.

10. The high-precision face recognition method for natural scenes applied to golf courses as claimed in claim 1, characterized in that: In step S45, the extracted feature vector is compared with known face features in the database, and the similarity between the input face and each face in the database is calculated using the cosine similarity method, and the result with the highest similarity is selected as the recognition result.

Images