A target detection method based on polarization dispersion degree

Abstract

The application provides a target detection method based on polarization discreteness, including the following contents: 1, constructing a target detection data set based on a polarization image; 2, deducing a physical model of surface polarization discreteness of a target object; 3, designing a network structure for extracting polarization discreteness based on the polarization discreteness model; 4, designing a target detection model suitable for the polarization image in combination with a target detection framework based on deep learning; 5, optimizing model parameters and improving model generalization ability. Compared with a traditional target detection model suitable for a light intensity image, the application designs a brand-new network structure based on an optical model for extracting the discreteness features of the polarization image, suppresses the influence of a scattering medium, utilizes the difference in polarization characteristics of different target objects, and realizes accurate detection of the target object in a complex and fuzzy environment, thereby providing a new solution for the target detection field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of polarization imaging technology, and in particular to a target detection method based on polarization dispersion. BACKGROUND

[0002] In complex natural environments, there are all kinds of scattering media, which leads to poor image quality of light intensity obtained by traditional optical imaging, and brings great challenges to target detection tasks. With the breakthrough of deep learning technology, the current mainstream target detection scheme is to use convolutional neural network, which has the characteristics of feature learning through data-driven. Compared with the manual design of features according to prior knowledge in traditional methods, convolutional neural network can automatically express more complex mathematical models and has stronger generalization ability. But with the application of target detection technology in actual production and life, the detection accuracy of the existing target detection technology in the environment full of scattering media is greatly affected. The current solution is mainly to improve the network structure to extract as much image information as possible, but there is still a lack of solution to realize accurate target detection in high-concentration scattering media.

[0003] Polarization imaging technology is widely used in the detection of specific target objects in pure environments. The application of polarization imaging technology to target detection technology has attracted widespread attention. However, in these studies, polarization information is mainly used to detect the material structure of the target object or to improve the contrast between the target object and the background information. The related research on the combination of target detection technology is more to replace the light intensity image with the polarization image input into the target detection model. However, without considering the polarization principle and without constructing a data set suitable for the polarization image target detection model, directly inputting the polarization image into the target detection network will cause the target detection network to fail to fully utilize the polarization information, which makes it extremely difficult to improve the precision of the target detection model. It has great application value to establish a polarization optical physics model to improve the precision of the target detection model. SUMMARY

[0004] In order to overcome the defects in the prior art, the present application provides a target detection method based on polarization dispersion, establishes a polarization dispersion model, constructs a network module to extract the dispersion features of the polarization image, and realizes the detection of the target object. The polarization dispersion features of the target object are extracted from the high-concentration scattering medium environment to realize accurate detection of the target object.

[0005] TECHNICAL SCHEME

[0006] A target detection method based on polarization dispersion, comprising the following steps:

[0007] Step one, construct the target detection dataset of polarization image, and pretreat the target detection dataset, group according to the scattering turbidity of the scattering medium in the image, then label the images in the target detection dataset, mark the position and category of the image target, and construct the original data suitable for training the target detection model;

[0008] Step two, combine the polarization bidirectional reflectance distribution function (pBRDF) model and the Fresnel reflection model to derive the distribution rule of the reflected light polarization information of the target surface, and derive the relationship expression between the physical properties of the target surface and the polarization information of the reflected light of the target surface;

[0009] Step three, the physical properties of target surfaces of different materials are different, and the polarization imaging technology can store this difference into the polarization image. However, in an environment full of high-concentration scattering medium, a large amount of noise inevitably appears in the polarization image, which interferes with the extraction of polarization features. According to the pixel difference between the target region and the background region, based on the statistical idea, by statistically analyzing the change of polarization information in the polarization image, the polarization information submerged in the noise can be highlighted. According to the measurement method of dispersion in statistics, the expression for calculating the dispersion of the polarization image is derived, and the polarization dispersion of the target is given by the following formula:

[0010]

[0011] Based on the convolutional neural network structure, the network layer for extracting the polarization dispersion feature is constructed according to the expression;

[0012] Step four, combine the mainstream target detection network framework YOLOv4 to construct the target detection network model based on polarization dispersion, and the target detection network model is composed of polarization state calculation layer, dispersion calculation network layer and target detection network.

[0013] Step five, the polarized component images (0°, 45°, 90° linear polarization) are taken as the input of the model, the polarized component images are subjected to matrix operation according to the Stokes vector in the polarization state calculation layer, the Stokes parameters are calculated, and then the AOP and DOLP images are obtained, and then the AOP and DOLP images are transmitted into the dispersion calculation network layer constructed, the dispersion of the polarization characteristics is calculated using the pixel gray value of the AOP and DOLP images, the distribution state of the polarization information is obtained, the polarization dispersion image is convolved, the image size meets the input size requirement of the backbone network, the physical characteristic parameters of the target object are represented by higher-dimensional characteristic information in the backbone network, different dimensions of feature maps are processed in the neck network, more dimensions of features are obtained by combining global information and local information, and then the feature maps are modified, and then the position and category information of the target object are output in the head network according to the probability distribution, the boundary box is drawn on the original light intensity image according to the target object position coordinates and category information, and the text information of the category and confidence of the target object is labeled.

[0014] Step six, the network is trained using the target detection data set in step one, the weight file of the target detection model is obtained, and the target detection network model in step four is tested using the test set image, and the Mean Average Precision and Intersection Over Union evaluation indexes are used to measure the model effect.

[0015] Further, the construction method of the target detection data set in step one: a common scattering medium is used to create a complex environment in air or water, objects of different materials are placed in the environment as target objects, based on the active polarization imaging principle, the polarized light is modulated as an active light source, and the scattering medium concentration is controlled, the polarization component images of the target objects in different environments are photographed using a CMOS camera equipped with a polarizer, the target detection data set of the original polarization image is obtained, the polarization information in the target detection data set of the original polarization image is extracted, the multi-dimensional polarization information such as the polarization angle and the degree of polarization in the image is obtained, and the data disturbed by noise is removed, and the data is cropped and transposed based on data screening to improve the quality of the target detection data set.

[0016] Further, the scattering medium includes smoke and skimmed milk.

[0017] Further, the device used in the construction method of the target detection data set in step one includes a light source, a polarizer, a quarter-wave plate and a beam expander are sequentially arranged on one side of the light source, a glass water tank is arranged on the side of the beam expander away from the quarter-wave plate, the glass water tank contains a scattering medium, the scattering medium contains target objects, and a CMOS camera and a stepping motor equipped with a polarizer are further arranged on the side of the glass water tank close to the beam expander.

[0018] Advantages

[0019] Compared with the prior art, the present application has the following advantages: a polarization dispersion model is established, a network module is constructed to extract the dispersion characteristics of the polarization image to realize detection of the target object, the polarization dispersion characteristics of the target object are extracted from a high-concentration scattering medium environment to realize accurate detection of the target object, and further use of polarization information is made on the basis of a target detection model based on deep learning to detect the target object in a complex environment, thereby providing a new solution for the target detection field. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 An overall flowchart of the target detection method based on polarization dispersion is shown in the figure.

[0021] Figure 2 A structural schematic diagram of the device used in the construction method of the target detection data set is shown in the figure.

[0022] Figure 3 A schematic diagram of the established polarization dispersion target detection network is shown in the figure.

[0023] Figure 4 A detection effect diagram of the present application is shown in the figure.

[0024] Figure 5 A comparison result table of the detection effect of the present application is shown in the figure.

[0025] Reference Signs:

[0026] Light source 1, polarizer 2, quarter-wave plate 3, beam expander 4, target object 5, glass water tank 6, scattering medium 7, CMOS camera 8, step motor 9 equipped with a polarizer. DETAILED DESCRIPTION

[0027] To better illustrate the content of the present application, the following will be described in conjunction with the figures and examples:

[0028] There are Figures 1-5 As shown in the figure, the present application discloses a target detection method based on polarization dispersion, which comprises the following steps:

[0029] Step 1, construct a target detection data set of polarization images, and pretreat the target detection data set, group according to the scattering turbidity of the scattering medium in the image, then label the images in the target detection data set, mark the position and category of the image target object, and construct original data suitable for training a target detection model;

[0030] Step two, combine the polarized bidirectional reflectance distribution function (pBRDF) model and the Fresnel reflection model to derive the distribution rule of the reflected light polarization information of the target surface, and derive the relationship expression between the physical properties of the target surface and the polarization information of the reflected light of the target surface;

[0031] Step three, the physical properties of different material target surfaces are different, and the polarization imaging technology can store this difference into the polarization image, however, in the environment full of high concentration scattering medium, a large amount of noise inevitably appears in the polarization image, which interferes with the extraction of polarization features, according to the pixel difference between the target region and the background region, based on the statistical idea, by counting the change of polarization information in the polarization image, the polarization information submerged in the noise can be highlighted, according to the measurement method of dispersion degree in statistics, the expression for calculating the dispersion degree of the polarization image is derived, and the polarization dispersion degree of the target is given by the following formula:

[0032]

[0033] Based on the convolution neural network structure, the network layer for extracting the polarization dispersion degree feature is constructed according to the expression;

[0034] Step four, combine the mainstream target detection network framework YOLOv4 to construct a target detection network model based on polarization dispersion degree, and the target detection network model is composed of polarization state calculation layer (PSC), dispersion degree calculation network layer (DC) and target detection network (YOLOv4).

[0035] Step five, the polarization component image (0°, 45°, 90° linear polarization) is taken as the input of the model, the polarization component image is operated by matrix according to the Stokes vector in the polarization state calculation layer, the Stokes parameters are calculated, and the AOP and DOLP images are obtained, then the AOP and DOLP images are input into the dispersion degree calculation network layer constructed, the dispersion degree of the polarization characteristics is calculated using the pixel gray value of the AOP and DOLP images, the distribution state of the polarization information is obtained, the polarization dispersion degree image is convolved, the image size meets the input size requirement of the backbone network, in the backbone network, the physical characteristic parameters of the target are represented by higher dimensional characteristic information, in the neck network, different dimensional feature maps are processed, more dimensional features are obtained by combining global information and local information, and the feature map is modified, then according to the probability distribution, the position and category information of the target are output in the head network, the boundary box is drawn on the original light intensity image according to the position coordinates and category information of the target, and the text information of the category and confidence of the target is labeled.

[0036] Step six, the target detection data set in step one is used for training the network built to obtain a weight file of a target detection model, and the target detection network model in step four is tested by using a test set image, and Mean Average Precision (mAP) and Intersection Over Union (IOU) evaluation indexes are used to measure the model effect.

[0037] Further, the construction method of the target detection data set in step one is as follows: a common scattering medium 7 is used to create a complex environment in air or water, objects with different materials are placed in the environment as target objects 5, based on the active polarization imaging principle, a polarized light is modulated as an active light source, the concentration of the scattering medium is controlled, a CMOS camera 8 equipped with a polarizer is used to shoot the polarization component images of the target objects 5 in different environments, the target detection data set of the original polarization images is obtained, the polarization information in the target detection data set of the original polarization images is extracted, the multi-dimensional polarization information such as the polarization angle and the degree of polarization in the images is obtained, the data disturbed by noise is removed, and data cropping and transposition processing is performed on the basis of data screening to improve the quality of the target detection data set.

[0038] Further, the scattering medium includes smoke and skimmed milk.

[0039] Further, the device used in the construction method of the target detection data set in step one includes a light source 1, a polarizer 2, a quarter-wave plate 3 and a beam expander 4 are sequentially arranged on one side of the light source 1, a glass water tank 6 is arranged on the side of the beam expander 4 away from the quarter-wave plate 3, a scattering medium 7 is arranged in the glass water tank 6, a target object 5 is arranged in the scattering medium 7, and a CMOS camera 8 and a stepping motor 9 equipped with a polarizer are further arranged on the side of the glass water tank 6 close to the beam expander 4.

[0040] In order to quantitatively evaluate the detection effect, Mean Average Precision (mAP) and Intersection Over Union (IOU) evaluation indexes are used to measure the model effect, and it can be directly seen from the experimental results that the application can realize accurate detection of the target object in a high-concentration scattering medium environment. Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the application, and not to limit them; although the technical solutions of the application have been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for constructing a target detection dataset based on polarization images, characterized in that, Includes the following steps: Step 1: Construct a target detection dataset of polarized images and preprocess the dataset. Group the images according to the scattering turbidity of the scattering medium in the images. Then label the images in the target detection dataset, marking the location and category of the target objects in the images, and construct the original data suitable for training the target detection model. Step 2: Combining the polarization bidirectional reflection distribution function (pBRDF) model and the Fresnel reflection model, derive the distribution law of the polarization information of the reflected light from the target surface, and derive the relationship expression between the physical properties of the target surface and the polarization information of the reflected light from the target surface. Step 3: Different target surfaces have different physical properties. Polarization imaging technology can store these differences in polarization images. However, in environments filled with high-concentration scattering media, polarization images inevitably contain a lot of noise, which interferes with the extraction of polarization features. Based on the pixel differences between the target area and the background area, and using statistical methods, the polarization information submerged in noise can be highlighted by statistically analyzing the changes in polarization information in the polarization image. Based on the method of measuring dispersion in statistics, an expression for calculating the dispersion of polarization images is derived. The polarization dispersion of the target object is given by the following formula: Based on the construction of a convolutional neural network, a network layer for extracting polarization discreteness features is constructed according to the expression. Step 4: Combine the mainstream target detection network framework—YOLOv4, and construct a target detection network model based on polarization discreteness. The target detection network model consists of a polarization state calculation layer, a discreteness calculation network layer, and a target detection network. Step 5: Using the polarization component images (0°, 45°, 90° linear polarization) as input to the model, the polarization state calculation layer performs matrix operations on the polarization component images based on the Stokes vector to calculate the Stokes parameters, thereby obtaining the AOP and DOLP images. Then, the AOP and DOLP images are fed into the constructed discreteness calculation network layer. The pixel gray values ​​of the AOP and DOLP images are used to calculate the discreteness of the polarization characteristics to obtain the distribution state of polarization information. In the backbone network, higher-dimensional characteristic information is used to characterize the physical characteristic parameters of the target object. The neck network processes feature maps of different dimensions, combines global and local information to obtain more dimensional features, and modifies the feature maps. Then, based on the probability distribution, the head network outputs the position and category information of the target object. Based on the position coordinates and category information of the target object, a bounding box is drawn on the original light intensity image, and the category and confidence text information of the target object are labeled. Step 6: Train the network using the object detection dataset described in Step 1 to obtain the weight file of the object detection model. Test the object detection network model described in Step 4 using test set images and use Mean Average Precision and Intersection Over Union evaluation metrics to measure the model performance.

2. The method for constructing a target detection dataset based on polarization images according to claim 1, characterized in that, The method for constructing the target detection dataset in step one is as follows: A complex environment is created in air or water using a common scattering medium 7. Objects of different materials are placed in this environment as target objects 5. Based on the principle of active polarization imaging, polarized light is modulated as the active light source, and the concentration of the scattering medium is controlled. A CMOS camera 8 equipped with a polarizer is used to capture polarization component images of the target objects 5 in different environments, obtaining the target detection dataset of the original polarization images. Polarization information is extracted from the target detection dataset of the original polarization images, and multi-dimensional polarization information such as polarization angle and polarization degree is obtained. Simultaneously, data interfered with by noise is removed. Based on data filtering, data cropping and transposition processing is performed to improve the quality of the target detection dataset.

3. The method for constructing a target detection dataset based on polarization images according to claim 2, characterized in that, The device used in the method for constructing the target detection dataset in step one includes a light source (1). A polarizer (2), a quarter-wave plate (3), and a beam expander (4) are arranged sequentially on one side of the light source (1). A glass water tank (6) is arranged on the side of the beam expander (4) away from the quarter-wave plate (3). A scattering medium (7) is placed inside the glass water tank (6). A target object (5) is placed inside the scattering medium (7). A CMOS camera (8) and a stepper motor (9) equipped with a polarizer are also arranged on the side of the glass water tank (6) near the beam expander (4).

4. The method for constructing a target detection dataset based on polarization images according to claim 3, characterized in that, The scattering medium includes smoke and skim milk.

Images