A method for detecting iris roll wheel based on double-flow structure

Abstract

The application discloses a pupil contraction wheel detection method based on a double-flow structure and relates to an eye iris detection method.The technical scheme of the application comprises iris image normalization, pupil contraction wheel boundary line detection, pupil contraction wheel fuzzy curve detection, determination of a detection area and area gradient detection (a preprocessing module, a texture flow module, a structure flow module, a fusion module and a detection module). In the iris normalized image, a double-flow network structure is used to extract structure features and texture features respectively. An edge gradient operator is used to detect in the initial positioning area to extract the pupil contraction wheel boundary. The application further reduces the pupil contraction wheel boundary fluctuation on the basis of ensuring the detection rate and the detection speed. Meanwhile, the application overcomes the unfavorable factors such as eyelashes, light spots and eyelids in the pupil contraction wheel detection process. The method is applied to an iris recognition system and a human body health monitoring system based on the iris.

Description

TECHNICAL FIELD

[0001] The application relates to an eye iris detection method, in particular to an eye iris contraction wheel detection method based on a double-flow structure. BACKGROUND

[0002] The iris contraction wheel is the boundary between the ciliary part and the pupil part in the iris. The position distribution of the iris contraction wheel, as one of main features of the iris, can be used as an important basis for iris recognition and iris diagnosis. For extraction of the contraction wheel, due to the complex texture distribution near the contraction wheel and the existence of interference such as a light spot, an eyelid and eyelashes, the existing extraction methods are difficult to exclude the influence of the internal and external textures and interference factors of the contraction wheel.

[0003] At present, the iris contraction wheel detection methods mainly include a gradient extremum-based method, a structure pattern-based method, a Snake operator-based method, an iris texture-based method, an ant colony algorithm-based method and a genetic algorithm-based method. The above methods have the problems of poor anti-interference performance of the contraction wheel detection result, unsatisfactory detection precision and low time efficiency. The method can ensure that the selected preliminary positioning region contains the features of the contraction wheel and reduces the interference of the light spot, the eyelashes, the eyelid and the internal texture of the contraction wheel in the maximum range, is beneficial to detection by a subsequent edge gradient operator, and further reduces the boundary fluctuation of the contraction wheel on the basis of ensuring the detection rate and the detection speed.

[0004] An application patent with the Chinese patent number CN201410258004.X discloses an iris pupil-to-contraction wheel region information extraction method. The method is realized through the steps of human eye image acquisition, human eye iris positioning, iris image normalization, establishment of a basic element structure and a pattern, establishment of an image feature vector, definition of an image boundary and a non-boundary pattern, and extraction of the outer edge of the pupil-to-contraction wheel region. However, due to the similarity between the iris pit texture boundary and the contraction wheel boundary, it is difficult to position the contraction wheel boundary for an iris image with many pits, and the edge width and the anti-noise ability are difficult to be considered.

[0005] An application patent with the Chinese patent number CN201510924372.8 discloses a method for extracting a contraction wheel based on a two-stage ant colony algorithm. The method needs to be iterated twice. The task of the first stage is to determine the approximate position of the ANW by using a large number of artificial ants for a small number of iterations, and the position is used to determine the initial position of the artificial ants in the next stage. The second stage uses a small number of artificial ants for multiple iterations to obtain the final contraction wheel extraction result. Although the accuracy is improved, the detection speed is slow.

[0006] It can be seen that the iris detection in the field has the problems of light spot, eyelash and eyelid interference and iris contraction wheel positioning fluctuation. SUMMARY

[0007] The purpose of this invention is to provide a method for detecting iris roll-off wheels based on a dual-flow structure. This method includes iris image normalization, roll-off wheel boundary line detection, roll-off wheel blur curve detection, determination of the detection region, and region gradient detection (preprocessing module, texture flow module, structure flow module, fusion module, and detection module). This invention further reduces boundary fluctuations in iris roll-off wheel detection while maintaining detection rate and speed. This method can be applied to iris recognition systems and iris-based human health monitoring systems.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A method for detecting iris retraction rings based on a dual-flow structure, the method comprising the following technical solutions:

[0010] Image normalization, which uses the rubber sheet model proposed by Daugman ( Figure 1 Convert the iris ciliary region into a rectangular image of 200 rows * 720 columns (pixels). Figure 2 ); Coil winding boundary line detection: its V-type operator (such as Figure 3 The texture inside the scroll wheel is obtained, and the bottom boundary of the texture inside the scroll wheel is located using a custom density ratio index (Equation 1); Scroll wheel blur curve detection: it constructs an improved ResSegNet model by adding residual learning to the network ( Figure 7 In the encoder part, conv1 uses 7 convolutional layers with 64 kernels for shallow feature extraction; conv2~4 use double convolutional layers with 128, 256 and 512 kernels respectively, stacked 4, 6 and 6 times respectively, with a max pooling layer stride of 2, and two residual blocks are used; in the decoder part, the symmetry between the encoder and decoder is maintained, the decoder contains two residual blocks and the number of convolutional kernels in each layer of the residual blocks corresponds to that in the encoder; at the same time, it includes 4 upsampling layers corresponding to the 4 pooling layers in the encoder, with an upsampling layer stride of 2; in the model output part, n convolutional layers with a kernel size of 1 are used, where n is the number of target categories to be classified in the iris image, the probability value of each pixel in the image is calculated through the Softmax layer, and finally the semantic segmentation result of the image is output through the output layer; the detection region is determined, and its feature fusion module (Equation 2) analyzes the blur curve of the convolution wheel and the boundary line of the convolution wheel to obtain the detection region; the region gradient detection uses the Scharr operator ( Figure 9 ) Detect extreme points in the vertical direction, connect all extreme points, and obtain the final positioning of the coiling wheel boundary; This is used in conjunction with an iris image library ( Figure 10 It uses the HM9918 handheld iris scanner ( Figure 11 As a data acquisition device, it collected a visible light iris image library containing various iris textures and interference factors.

[0011] The V-shaped operator is used for detecting the internal texture of the iris roll-up wheel (S2) Figure 4 The definition is realized by analyzing the gray scale and shape features of the internal texture.

[0012] The density ratio index is used to determine the roll-up wheel boundary line (S3) Figure 5 The definition is realized by statistical principles and the internal texture detection result (S2) Figure 6

[0013] The method is to prevent gradient explosion by adding a BN layer for batch normalization after each convolutional layer; in order to prevent network degradation and gradient disappearance, a residual learning module (S4) is introduced every two convolutional layers. Figure 8

[0014] The advantages and effects of the present application are as follows:

[0015] The method includes iris image normalization, roll-up wheel boundary line detection, roll-up wheel fuzzy curve detection, determination of the detection area and area gradient detection (preprocessing module, texture flow module, structure flow module, fusion module and detection module); the present application further reduces the iris roll-up wheel detection method with boundary fluctuation on the basis of ensuring the detection rate and detection speed. The method can be applied to the iris recognition system and the human health monitoring system based on the iris. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a schematic diagram of the overall structure scheme of the present application;

[0017] Figure 2 is a schematic diagram of the iris ciliary region normalization result of the present application.

[0018] Figure 3 is a schematic diagram of the V-shaped detection operator of the present application;

[0019] Figure 4 is an enlarged view of the internal texture of the iris roll-up wheel of the present application;

[0020] Figure 5 is a schematic diagram of the roll-up wheel boundary line detection result of the present application;

[0021] Figure 6 is a schematic diagram of the internal texture detection result of the roll-up wheel of the present application;

[0022] Figure 7 is a schematic diagram of the improved ResSegNet model of the present application;

[0023] ​​Figure 8 is a schematic diagram of the improved ResSegNet model residual learning module of the present application;

[0024] Figure 9 is a schematic diagram of the vertical direction Scharr operator of the present application;

[0025] Figure 10 is a schematic diagram of the matching iris library of the present application;

[0026] Figure 11 is a schematic diagram of the HM9918 type handheld iris instrument used in the present application. DETAILED DESCRIPTION

[0027] The present application will be described in detail below in conjunction with the embodiments shown in the drawings.

[0028] An iris roll-up wheel detection method based on a double-flow structure, comprising iris image normalization, roll-up wheel boundary line detection, roll-up wheel fuzzy curve detection, determination of a detection region and region gradient detection (a preprocessing module, a texture flow module, a structure flow module, a fusion module and a detection module).

[0029] I. Iris image normalization

[0030] The rubber sheet model proposed by Daugman is adopted to change the annular region into a rectangular image with the same size of 200*720 columns. Since the size of the iris varies from person to person, the normalization method can normalize the iris regions of different sizes into the same rectangular image. When the iris is unfolded, the radial direction of the iris is taken as the row direction of the rectangle, and the circumferential direction of the iris is taken as the column direction of the unfolded rectangular image.

[0031] II. Roll-up wheel boundary line detection

[0032] The V-shaped detection operator is used to obtain the corresponding roll-up wheel internal texture, and the self-defined density ratio index is used to locate the bottom boundary line of the roll-up wheel internal texture. The number of texture pixel points of all rows is accumulated to obtain the value of the density ratio of the first row , and then the density ratio corresponding to all rows is obtained Therefore, the row corresponding to the maximum value of the density ratio is the bottom boundary line of the roll-up wheel internal texture.

[0033] III. Roll-up wheel fuzzy curve detection

[0034] A 50-layer ResSegNet network is adopted. In order to prevent gradient explosion, a BN layer is added for batch normalization after each convolution layer; in order to ensure the nonlinear mapping, the convolution layer is composed of two layers and an activation function exists; in order to prevent network degradation and gradient disappearance, a residual learning module is introduced every two convolution layers. ​

[0035] In the encoder section, conv1 uses 7 convolutional layers with 64 kernels each for shallow feature extraction; conv2-4 use double convolutional layers with 128, 256, and 512 kernels respectively, stacked 4, 6, and 6 times, with a max pooling stride of 2, and two residual blocks. In the decoder section, maintaining the symmetry between the encoder and decoder, the decoder contains two residual blocks, and the number of convolutional kernels in each layer of the residual blocks corresponds to that in the encoder. It also includes 4 upsampling layers corresponding to the 4 pooling layers in the encoder, with an upsampling stride of 2. In the model output section, n convolutional layers with a kernel size of 1 are used, where n is the number of target categories to be classified in the iris image. Then, a Softmax layer calculates the probability value of each pixel in the image, and finally, the output layer outputs the localization result of the convolution wheel blur curve.

[0036] IV. Determining the Testing Area

[0037] Through The detection area of ​​the iris rollup wheel is obtained by fusing the fuzzy positioning curve of the iris rollup wheel and the boundary line between the inner and outer textures of the rollup wheel.

[0038] V. Regional Gradient Detection

[0039] Because there is a lot of noise in the detection area of ​​the coiled wheel, it needs to be denoised by Gaussian filtering before detection. Then, the Scharr gradient operator is used to detect extreme points in the vertical direction in the filtered area, and all extreme points are connected to obtain the final location of the coiled wheel boundary. Example

[0040] 1) First, normalize the iris image, then input the normalized iris image. Figure 1 Semantic segmentation was performed on the mid-structure flow backbone network to obtain the fuzzy localization curve of the iris roll-up wheel.

[0041] 2) Input the grayscale normalized iris unfolded image. Figure 1 The texture features inside the iris scroll wheel are extracted using a texture flow network. The texture inside the scroll wheel is extracted using a V-shaped detection operator. The total number of texture pixels corresponding to each row is calculated. Finally, a custom density ratio metric is used to find its maximum value; the corresponding horizontal coordinate value is the boundary between the texture inside and outside the scroll wheel.

[0042] 3) Combine the positioning curve of the scroll wheel obtained from the structure flow and the boundary line between the inner and outer textures of the scroll wheel obtained from the texture flow to determine the initial search area of ​​the scroll wheel. If the area contains interference factors, replace the pixel value of the area with the gray average value of the surrounding pixels.

[0043] 4) Gaussian filtering is performed in the initial search range of the pucker wheel to remove noise; Scharr gradient operator is used to detect in the vertical direction, and a gradient matrix with the same number of pixels as the detection area is generated after operation. The extreme points in the column direction are found in the gradient matrix, and the obtained gradient extreme points are connected on the original image to obtain the final positioning of the pucker wheel boundary.

[0044] The above describes in detail the iris pucker wheel detection method based on the double-flow structure provided by the present application. The principles and implementation modes of the present application are described by applying specific examples in this paper. The above examples are only used to help understand the method of the present application and its core idea. It should be pointed out that, for those skilled in the art, without departing from the principles of the present application, the present application can be improved in several ways, and these improvements also fall within the protection scope of the claims of the present application.

[0045] Formula 1: Density ratio index

[0046] wherein , , respectively, correspond to the height and width of the normalized iris image; is the pixel value of the image after texture detection and binarization, represents is the texture pixel.

[0047] Formula 2: Fusion formula

[0048]

[0049] wherein, is the positioning area, represents the horizontal coordinate value corresponding to the horizontal dividing line at the bottom of the pucker wheel inside texture obtained by the texture flow, represents the horizontal coordinate value of the first column pucker wheel positioning initial curve obtained by the structure flow. The parameter 5 in formula (4) is used to exclude the influence of texture flow dividing line error.

Claims

1. A method for detecting the iris roll wheel of an eye based on a double flow structure, characterized in that, The method comprises the following schemes: Image normalization, which converts the iris ciliary region into a rectangular image of 200 rows * 720 columns pixels using the rubber sheet model proposed by Daugman; Collapsible wheel boundary detection: the V-shaped operator obtains the internal texture of the collapsible wheel, and the bottom boundary of the internal texture is located by using a self-defined density ratio index; Collapsible wheel fuzzy curve detection: an improved ResSegNet model is constructed by adding residual learning in the network; in the encoder part, conv1 adopts a convolution layer with 7 convolution kernels and a number of 64 to extract shallow features; conv2~4 respectively adopt double convolution layers with convolution kernel numbers of 128, 256 and 512, and are respectively stacked for 4, 6 and 6 times, the step length of the maximum pooling layer is 2, and two residual blocks are used; in the decoder part, the symmetry of the encoder and the decoder is maintained, the decoder includes two residual blocks and the convolution kernel numbers of each layer in the residual block are corresponding to those in the encoder; at the same time, four up-sampling layers corresponding to the four pooling layers in the encoder are included, and the step length of the up-sampling layer is 2; The model output part uses n convolution layers with a convolution kernel size of 1, n is the number of target categories to be classified of the iris image, the probability values of each pixel of the image are calculated through a Softmax layer, and finally the semantic segmentation result of the image is output through an output layer; the detection region is determined, The feature fusion module analyzes the collapsible wheel fuzzy curve and the collapsible wheel boundary to obtain the detection region; the region gradient detection utilizes a Scharr operator to detect extreme points in the vertical direction, connects all the extreme points to obtain the final positioning of the collapsible wheel boundary; and the iris image library is used in conjunction therewith, the HM9918 type handheld iris instrument is used as the acquisition equipment to collect a visible light iris image library containing various iris textures and interference factors. The density ratio index is used to determine the boundary line of the scroll wheel, and is defined based on statistical principles and internal texture detection results. The internal texture of the scroll wheel that meets the requirements is obtained through a V-shaped detection operator, and the bottom boundary line of the internal texture of the scroll wheel is located using a custom density ratio index. The number of texture pixels in all rows is accumulated to obtain the... row density ratio value This allows us to obtain the density ratios corresponding to all rows. The row corresponding to the maximum density ratio is the bottom boundary line of the texture inside the shrink wheel.

2. The method for detecting the iris roll wheel based on the double-flow structure according to claim 1, characterized in that, The V-shaped operator is used to detect the internal texture of the iris collapsible wheel, and the definition is realized by analyzing the gray scale and shape features of the internal texture.

3. The method of claim 1, wherein the method is based on a double-flow structure. The method is used to prevent gradient explosion, and a BN layer is added after each convolution layer for batch normalization; in order to prevent network degradation and gradient disappearance, a residual learning module is introduced every two convolution layers.

Images