Visual biomimetic edge detection method based on texture gradient adjustment

Abstract

This invention provides a visual biomimetic edge detection method based on texture gradient modulation, belonging to the field of image edge detection technology. It focuses on introducing texture gradients to detect significant edges and suppress texture. First, the retina is modeled, and the image is encoded to obtain multiple information channels. Based on this, a novel peripheral modulation mechanism is designed, including unidirectional facilitation, lateral inhibition, and omnidirectional inhibition, to modulate the responses of simple cells in the primary visual cortex. Simultaneously, texture gradients are extracted using texture information and combined with the responses of simple cells to highlight the boundaries of textured regions and weaken the responses at textured edges. Next, endpoint cells in the second-level visual cortex are modeled to further modulate edge responses. Finally, the fourth-level visual cortex is modeled to integrate edge cues from all information channels to obtain the final edge detection result. This invention solves the problems of poorly highlighted textured region boundaries and significant noise in existing edge detection methods.

Description

Technical Field

[0001] This invention belongs to the field of image edge detection technology, and particularly relates to a visual biomimetic edge detection method based on texture gradient adjustment. Background Technology

[0002] Visual neuroscience research has shown that contour information is crucial for object recognition, and it is even possible to identify objects solely based on their contours. Existing biomimetic edge detection methods have achieved good results by establishing hierarchical parallel frameworks and introducing non-classical receptive field peripheral modulation mechanisms to modulate edge responses. However, these methods do not consider how to utilize texture information to highlight the boundaries of textured regions and reduce texture responses. Consequently, the final edge detection results either contain significant noise or exhibit weak or unclear hierarchical distinctions. Therefore, it is necessary to research methods based on visual biomimetic edge detection. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a visual biomimetic edge detection method based on texture gradient adjustment. This method focuses on introducing texture information to detect significant edges, solving the problems of indistinct texture region boundaries and obvious noise in existing edge detection methods.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] This solution provides a visual biomimetic edge detection method based on texture gradient adjustment, including the following steps:

[0006] S1. Obtain the edge detection image and construct the retinal-coded color, brightness, and brightness contrast channels to obtain multi-channel information;

[0007] S2. Based on the multi-channel information, construct the classical receptive field of simple cells in the primary visual cortex using the first-order gradient of a two-dimensional Gaussian function, and respond to local edges;

[0008] S3. Based on the response results, the edge response is adjusted using the surrounding adjustment mechanism, and the texture gradient is introduced to adjust the response of simple cells to obtain the edge response of the visual primary cortex.

[0009] S4. Based on the edge response of the primary visual cortex, construct endpoint cells and modulate them to obtain the edge response of the secondary visual cortex;

[0010] S5. Based on the edge response of the second-level visual cortex, the results of all channels are fused to obtain the edge response of the fourth-level visual cortex, thus completing the detection of visual bionic edges.

[0011] The beneficial effects of this invention are: This invention detects edges by mimicking the hierarchical perception of the human visual nervous system. By modeling the retina and lateral geniculate body, primary visual cortex, secondary visual cortex and quaternary visual cortex, it can effectively perform edge detection and solve the problems of non-prominent texture region boundaries and obvious noise in existing edge detection methods.

[0012] Further, step S1 includes the following steps:

[0013] S101. Obtain the edge detection image and preprocess the edge detection image using the square root function;

[0014] S102. Based on the preprocessing results, the color information of the edge detection image is encoded using the color antagonism mechanism of the retina;

[0015] S103. Encode the brightness information and brightness contrast information channels of the edge detection image, and obtain multi-channel information based on the encoding results of step S102.

[0016] The beneficial effects of the above-mentioned further solutions are as follows: When modeling the retina and lateral geniculate body, this invention utilizes the retina's perception and encoding of visual information and the lateral geniculate body's integration and transmission of the encoded visual information. First, a square root function is used to enhance image contrast; then, color information is encoded by mimicking the retina's color antagonism mechanism; finally, brightness and brightness contrast information are encoded. Ultimately, by performing a series of encodings on the image information, parallel color antagonism channels, brightness channels, and brightness contrast channels are obtained, which more closely approximates real human perception.

[0017] Furthermore, the preprocessing expression is as follows:

[0018]

[0019] Among them, C h I represents the red, green, and blue channel information of the edge detection image after preprocessing. h This represents the red, green, and blue channel information of the original edge detection image;

[0020] The expression for encoding the color information of the edge detection image is as follows;

[0021] Ch rg (x,y)=C r (x,y)-wC g (x,y)

[0022]

[0023] Where Ch represents the color antagonistic cell response, (x,y) represents the spatial location, w represents the color antagonistic weight, and C r (x,y) represents the preprocessed red channel, C g (x,y) represents the preprocessed green channel, Ch rg (x,y) represents the red-green color antagonistic channel, Ch by (x,y) represents the blue-yellow color antagonistic channel, C b (x,y) represents the preprocessed blue channel.

[0024] The beneficial effect of the above-mentioned further scheme is that it effectively encodes color channel information through a color antagonism mechanism for subsequent color edge detection.

[0025] Furthermore, the channel response of the brightness information is as follows:

[0026] Ch lu (x,y)=0.299C r (x,y)+0.587C g (x,y)+0.114C b (x,y)

[0027] The channel response for the brightness and contrast information is as follows:

[0028]

[0029] Avg(x,y)=A(x,y,r)*Ch lu (x,y)

[0030] A(x,y,r)=||Circle(x,y,r)||

[0031]

[0032] Among them, Ch lu (x,y) represents the brightness channel, Ch lc (x,y) represents the brightness / contrast channel, A(x,y,r) represents the circular average kernel, Avg(x,y) represents the convolution result of the brightness channel and the circular average kernel, Circle(x,y,r) represents a circle, ||·|| represents the normalization function, * represents the convolution operation, and r represents the radius of the local region of the circle.

[0033] Furthermore, the response expression of the classical receptive field is as follows:

[0034]

[0035]

[0036] Where CRF(x,y,θ) represents the response of the classical receptive field. Represents partial derivative, θ represents edge orientation, N θ Let i represent the total number of edge orientations, G(x,y,σ) represent the two-dimensional Gaussian function, σ represent the standard deviation of the two-dimensional Gaussian function, and Ch' represent the number of edge orientations. rg Ch' represents an unbalanced red-green antagonistic channel. by This indicates an unbalanced blue-yellow color antagonistic channel.

[0037] The beneficial effect of the above-mentioned further scheme is that by using the first derivative of the two-dimensional Gaussian function to model the classical receptive field of a simple cell, the local edge response of different orientations can be effectively obtained.

[0038] Furthermore, step S3 includes the following steps:

[0039] S301. Based on the response results, a model of the classical receptive field of simple cells is introduced to promote unidirectional synergy:

[0040] SF(x,y,θ)=E same (x,y,σ x ,σ y ,θ)*CRF(x,y,θ)

[0041]

[0042]

[0043] Where SF(x,y,θ) represents unidirectional promotion, and E same (·) represents the Gaussian function of an ellipse with edge orientation θ, (x,y) represents the spatial position, and σ x and σ y The shape of the classical receptive field is represented by a, b, and c, which are all coefficients in the elliptic Gaussian function.

[0044] S302. Introducing lateral inhibition to model the classical receptive field of simple cells:

[0045] LI(x,y,θ)=E lateral (σ x ,σ y ,θ)*CRF(x,y,θ)

[0046]

[0047] Where LI(x,y,θ) represents lateral inhibition, E lateral (·) indicates a lateral suppression kernel with an edge orientation of θ;

[0048] S303. Introducing peripheral inhibition to model the classic receptive field of simple cells:

[0049]

[0050] Where FI(x,y) represents the surrounding inhibition;

[0051] S304. Based on the modeling results of steps S301-S303, obtain the non-classical receptive field response:

[0052] nCRF(x,y,θ)=w1SF(x,y,θ)+w2LI(x,y,θ)+w3FI(x,y)

[0053] Where nCRF(x,y,θ) represents the nonclassical receptive field response, and w1, w2 and w3 represent the weights of unidirectional facilitation, lateral inhibition and peripheral inhibition, respectively;

[0054] S305, Introducing a texture kernel:

[0055] K(x,y,r,θ)=A top (x,y,r,θ)-A bottom (x,y,r,θ)

[0056] Where K(x,y,r,θ) represents the texture kernel, and A top and A bottom Let A(x,y,r) represent the upper and lower halves of the edge along θ, respectively.

[0057] S306. Detect texture region boundaries using multiple texture kernels of different sizes:

[0058]

[0059] x' = x - rcosθ, y' = y - rsinθ

[0060] Among them, T ms (x,y,θ) represents the result of convolving multiple texture kernels of different radius sizes with texture information, K(x',y',r,θ) represents the translated texture kernel, CRF(x',y',θ) represents the response of the classical receptive field after translation, and x' and y' represent the translation of (x,y) along the edge towards the θ direction. The distance after;

[0061] S307: By fusing non-classical receptive field responses with texture region boundaries and using nonlinear functions to modulate the responses of simple cells and texture gradients, the edge responses of the visual primary cortex are obtained.

[0062] E v1 (x,y)=SC max(x,y)·T max (x,y)

[0063]

[0064]

[0065]

[0066] Among them, E v1 (x,y) represents the response of the primary visual cortex, SC max (x,y) represents all simple cells This represents the maximum value operation, T. ms (x,y,θ) represents the sum of texture gradients of multiple radii with edge orientation θ, H(x”) represents a non-linear function, x” represents the variable of the non-linear function, and T max (x,y) represents the maximum value of all texture gradients.

[0067] The beneficial effects of the above-mentioned further solutions are as follows: This invention utilizes the first derivative of a two-dimensional Gaussian function to model the classical receptive field of a simple cell to obtain the local response toward the edge. However, the local edges detected by the classical receptive field of a simple cell are filled with noise and textured edges. This invention proposes a surrounding modulation mechanism composed of unidirectional promotion, lateral inhibition, and surrounding inhibition, based on the fact that the non-classical receptive field composed of the region surrounding the classical receptive field can effectively suppress edge noise and textured edges. At the same time, it uses texture information to calculate the texture gradient, which can effectively suppress textured edges and highlight the boundaries of textured regions.

[0068] Furthermore, step S4 includes the following steps:

[0069] S401. Based on the edge response of the primary visual cortex, the linear orientation nucleus is divided into left and right parts, and the cells at the left and right endpoints are modeled and denoted as K. l and K r The expression for the linear orientation towards the kernel is as follows:

[0070]

[0071] Where K(x,y,l,θ) represents the linear orientation toward the kernel, and l represents the length;

[0072] S402, using K l and K r Get the response of each location to the edge endpoint:

[0073] EI(x,y,θ)=|K l (l,θ)*E v1 (x,y)-K r(l,θ)*E v1 (x,y)|

[0074] Where EI(x,y,θ) represents the endpoint response value, K l (l,θ) represents the left-endpoint cell nucleus, K r (l,θ) represents the right-end nucleus;

[0075] S403. Based on the acquired endpoint responses, obtain the edge responses of the second-level visual cortex:

[0076]

[0077] Among them, E v2 (x,y) represents the response of the second-order visual cortex, k(l1,θ) represents a linear orientation kernel of length l1, and E v1 (x,y) represents the response of the primary visual cortex, c represents the weighting coefficient, and K(l2,θ) represents the linear orientation kernel of length l2.

[0078] The beneficial effects of the above-mentioned further scheme are: the present invention uses a simple linear orientation nucleus to model the response of the second visual cortex to the edge. Since the second visual cortex has a large number of endpoint cells, the endpoint cells respond to the endpoints but do not respond to the edges. By modeling the endpoint cells, the results of the edge response can be adjusted.

[0079] Furthermore, the expression for the edge response of the fourth-level visual cortex is as follows:

[0080]

[0081] Among them, E v4 (x,y) represents the edge response of the fourth level visual cortex.

[0082] The beneficial effect of the above-mentioned further scheme is that the edge detection results of multiple information channels are fused by modeling the fourth visual cortex to obtain the final edge. Attached Figure Description

[0083] Figure 1 This is a flowchart of the method of the present invention.

[0084] Figure 2 This is a schematic diagram illustrating the overall workflow of the edge detection method in this embodiment. Detailed Implementation

[0085] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0086] Example

[0087] like Figure 1 As shown, this invention provides a visual biomimetic edge detection method based on texture gradient adjustment, the implementation method of which is as follows:

[0088] S1. Obtain the edge detection image and construct retinal-coded color, brightness, and brightness contrast channels to obtain multi-channel information. The implementation method is as follows:

[0089] S101. Obtain the edge detection image and preprocess the edge detection image using the square root function;

[0090] S102. Based on the preprocessing results, the color information of the edge detection image is encoded using the color antagonism mechanism of the retina;

[0091] S103. Encode the brightness information and brightness contrast information channels of the edge detection image, and obtain multi-channel information based on the encoding results of step S102.

[0092] In this embodiment, to obtain the edge detection dataset image, firstly, three types of cone cells (red, green, and blue) are modeled, and the root square function is used to preprocess the input image to enhance image contrast, which can be expressed as:

[0093]

[0094] Among them, C h I represents the red, green, and blue channel information of the edge detection image after preprocessing. h This represents the red, green, and blue channel information of the original edge detection image.

[0095] Then, mimicking the color antagonism mechanism of the retina, color information is encoded, and the calculation method is as follows:

[0096] Ch rg (x,y)=C r (x,y)-wC g (x,y)

[0097]

[0098] Where Ch represents the color antagonistic cell response, (x,y) represents the spatial location, w represents the color antagonistic weight, and C r (x,y) represents the preprocessed red channel, C g (x,y) represents the preprocessed green channel, Ch rg (x,y) represents the red-green color antagonistic channel, Ch by (x,y) represents the blue-yellow color antagonistic channel, C b (x,y) represents the preprocessed blue channel. In this embodiment, w=1 is set to obtain the balanced color antagonistic channel {Ch}. rg ,Ch by}, setting w=0.5 to obtain the unbalanced color antagonistic channel {Ch' rg ,Ch' by}

[0099] In this embodiment, brightness and brightness contrast information are encoded. The response of the brightness information channel is shown below:

[0100] Ch lu (x,y)=0.299C r (x,y)+0.587C g (x,y)+0.114C b (x,y)

[0101] The channel response for the brightness and contrast information is as follows:

[0102]

[0103] Avg(x,y)=A(x,y,r)*Ch lu (x,y)

[0104] A(x,y,r)=||Circle(x,y,r)||

[0105]

[0106] Among them, Ch lu (x,y) represents the brightness channel, Ch lc (x,y) represents the brightness contrast channel, A(x,y,r) represents the circular average kernel, Avg(x,y) represents the convolution result of the brightness channel and the circular average kernel, Circle(x,y,r) represents a circle, ||·|| represents the normalization function, * represents the convolution operation, and r represents the local radius of the circle, which is set to 2.5. Finally, parallel color antagonism, brightness, and brightness contrast information channels are obtained.

[0107] In this embodiment, after the edge detection image is processed in step S1, it can simulate the retina to encode image information, including color, brightness and brightness contrast information.

[0108] S2. Based on the multi-channel information, construct the classical receptive field of simple cells in the primary visual cortex using the first-order gradient of a two-dimensional Gaussian function, and respond to local edges;

[0109] In this embodiment, the first-order gradient of a two-dimensional Gaussian function is used to model the response of the classical receptive field of a simple cell. The principle is to use the texture gradient to measure the response of the simple cell to texture suppression and to highlight the boundaries of the texture region. The calculation method of the response of the classical receptive field of a simple cell is as follows:

[0110]

[0111]

[0112] Where CRF(x,y,θ) represents the response of the classical receptive field. Represents partial derivative, θ represents edge orientation, N θ Let i represent the total number of edge orientations, G(x,y,σ) represent the two-dimensional Gaussian function, σ represent the standard deviation of the two-dimensional Gaussian function, and Ch' represent the number of edge orientations. rg Ch' represents an unbalanced red-green antagonistic channel. by This represents an unbalanced blue-yellow color antagonistic channel. In this embodiment, σ is set to 0.5, and N... θ Set it to 12.

[0113] S3. Based on the response results, the edge response is adjusted using the surrounding adjustment mechanism, and the texture gradient is introduced to adjust the response of simple cells to obtain the edge response of the visual primary cortex.

[0114] In this embodiment, since the local edges detected by S2 are filled with noise and textured edges, the non-classical receptive field composed of the region surrounding the classical receptive field can effectively suppress edge noise and texture. A novel surrounding regulation mechanism composed of unidirectional promotion, lateral inhibition, and omnidirectional inhibition is proposed for modeling. The implementation of the novel surrounding regulation mechanism is as follows:

[0115] S301. Based on the response results, a model of the classical receptive field of simple cells is introduced to promote unidirectional synergy:

[0116] SF(x,y,θ)=E same (x,y,σ x ,σ y ,θ)*CRF(x,y,θ)

[0117]

[0118]

[0119] Where SF(x,y,θ) represents unidirectional promotion, and E same (·) represents the Gaussian function of an ellipse with edge orientation θ, (x,y) represents the spatial position, and σ x and σ y This represents the shape of the classical receptive field, where a, b, and c are coefficients in the elliptic Gaussian function, and σ... x and σ y Their values ​​were set to 0.3 and 2.0 respectively.

[0120] S302. Introducing lateral inhibition to model the classical receptive field of simple cells:

[0121] LI(x,y,θ)=E lateral (σ x ,σ y ,θ)*CRF(x,y,θ)

[0122]

[0123] Where LI(x,y,θ) represents lateral inhibition, E lateral (·) indicates a lateral suppression kernel with an edge orientation of θ;

[0124] S303. Introducing peripheral inhibition to model the classic receptive field of simple cells:

[0125]

[0126] Where FI(x,y) represents the surrounding suppression, and r is the elliptic Gaussian kernel E same Half the size;

[0127] S304. Based on the modeling results of steps S301-S303, obtain the non-classical receptive field response:

[0128] nCRF(x,y,θ)=w1SF(x,y,θ)+w2LI(x,y,θ)+w3FI(x,y)

[0129] Where nCRF(x,y,θ) represents the nonclassical receptive field response, and w1, w2 and w3 represent the weights of unidirectional facilitation, lateral inhibition and peripheral inhibition, respectively;

[0130] S305. Introduce a texture kernel to detect the contour of textured regions and suppress texture:

[0131] K(x,y,r,θ)=A top (x,y,r,θ)-A bottom(x,y,r,θ)

[0132] Where K(x,y,r,θ) represents the texture kernel, and A top and A bottom Let A(x,y,r) represent the upper and lower halves of the edge along θ, respectively.

[0133] S306. Detect texture region boundaries using multiple texture kernels of different sizes;

[0134] In this embodiment, multiple texture kernels of different sizes are used to detect texture region boundaries, and these kernels are summed. The response of the texture edge is:

[0135]

[0136] x' = x - rcosθ, y' = y - rsinθ

[0137] Among them, T ms (x,y,θ) represents the result of convolving multiple texture kernels of different radius sizes with texture information, K(x',y',r,θ) represents the translated texture kernel, CRF(x',y',θ) represents the response of the classical receptive field after translation, and x' and y' represent the translation of (x,y) along the edge towards the θ direction. The distance after;

[0138] S307: By fusing non-classical receptive field responses with texture region boundaries and using nonlinear functions to adjust the responses of simple cells and texture gradients, the edge responses of the visual primary cortex are obtained.

[0139] In this embodiment, the edge and texture region boundaries detected by simple cells are finally fused, and the final edge response of the primary visual cortex is:

[0140] E v1 (x,y)=SC max (x,y)·T max (x,y)

[0141]

[0142]

[0143]

[0144] Among them, E v1 (x,y) represents the response of the primary visual cortex, SC max (x,y) represents all simple cells This represents the maximum value operation, T. ms(x,y,θ) represents the sum of texture gradients of multiple radii with edge orientation θ, H(x”) represents a non-linear function, x” represents the variable of the non-linear function, and T max (x,y) represents the maximum value of all texture gradients.

[0145] In this embodiment, the non-classical receptive field formed by the area surrounding the classical receptive field can effectively suppress the local edges full of noise and texture detected in step S2.

[0146] S4. Based on the edge response of the primary visual cortex, construct endpoint cells and modulate them to obtain the edge response of the secondary visual cortex;

[0147] In this embodiment, the second-level visual cortex contains a large number of endpoint cells, which respond to endpoints but not to edges. First, a simple linear orientation kernel is used to model its response to edges:

[0148]

[0149] Where K(x,y,l,θ) represents the linear orientation toward the kernel, and l represents the length;

[0150] Then, the linear orientation towards the nucleus is divided into left and right parts, and the cells at the left and right endpoints are modeled and denoted as K. l and K r In this embodiment, the linear oriented nucleus of length 3 is first divided into left and right parts to model the left and right endpoint cells, denoted as K. l and K r .

[0151] S402, using K l and K r Obtain the response of each location to the edge endpoint;

[0152] EI(x,y,θ)=|K l (l,θ)*E v1 (x,y)-K r (l,θ)*E v1 (x,y)|

[0153] Where EI(x,y,θ) represents the endpoint response value, K l (l,θ) represents the left-endpoint cell nucleus, K r (l,θ) represents the right-end nucleus;

[0154] S403. Based on the acquired endpoint responses, obtain the edge responses of the second-level visual cortex:

[0155]

[0156] Among them, E v2 (x,y) represents the response of the second-order visual cortex, k(l1,θ) represents a linear orientation kernel of length l1, and E v1 (x,y) represents the response of the primary visual cortex, c represents the weighting coefficient, K(l2,θ) represents the linear orientation kernel of length l2, l1=5, l2=3.

[0157] In this embodiment, step S4 can be used to model the endpoint cells of the second-level visual cortex and use them to adjust the edges detected by the primary visual cortex.

[0158] S5. Based on the edge response of the second-level visual cortex, the results of all channels are fused to obtain the edge response of the fourth-level visual cortex, thus completing the detection of visual bionic edges.

[0159] In this embodiment, in step S5, the fourth-level visual cortex is simply modeled, and the edge detection results of multiple information channels are fused by adding the results of all channels to obtain the final edge:

[0160]

[0161] Among them, E v4 (x,y) represents the edge response of the fourth level visual cortex.

[0162] In this embodiment, as Figure 2 As shown, Figure 2 The complete principle block diagram of the present invention based on biomimetic multi-level edge detection is given. Figure 2 In this context, "Input image" refers to the input image; "Retina and LGN" refers to the retina; "V1" represents the primary visual cortex; "V2" represents the secondary visual cortex; "V4" represents the quaternary visual cortex; "Surround modulation" refers to ambient modulation; "Texture gradient" refers to texture gradient; and "Endpoint cells" refers to endpoint cells. Figure 2 The first line is image preprocessing. After processing the original image using the root square function, the red, green, and blue color channels of the input edge detection image are obtained. The second line is color channel information encoding. By mimicking the color antagonism mechanism of the retina, parallel color antagonism, brightness, and brightness contrast information channels are obtained. The third line simulates the response of the primary visual cortex to image edges. First, a Gaussian gradient is used to simulate the response of the classical receptive field of simple cells in V1. To suppress texture edges, a novel surround modulation mechanism is applied, which includes unidirectional promotion, lateral inhibition, and omnidirectional inhibition. Texture information is used to calculate the texture gradient, which is used to suppress texture edges and highlight the boundaries of texture regions. Figure 2 The fourth line describes the modeling process of the second visual cortex. The second visual cortex contains a large number of endpoint cells, which respond to endpoints but not to edges. Therefore, the endpoint neurons in the second-level visual cortex are modeled, and they are used to modulate the edge response of the primary visual cortex. The fifth line models the fourth-level visual cortex by directly summing the results from all channels to fuse the edge cues of all information channels, obtaining the final output.

[0163] The effectiveness of the multi-level edge detection method in this scheme will be illustrated below with specific examples.

[0164] The edge detection method of this scheme was compared with various existing biomimetic edge detection methods on three datasets. On the BSDS500 dataset, standard evaluation metrics were selected, namely average accuracy and F-score as common evaluation methods. The method was compared with various existing edge detection methods, including simple methods such as Canny operator, and biomimetic methods such as Butterfly-shaped, PC / BC, CO, MCI, SCO, SED, MCH, and DIDY. The final comparison results are shown in Table 1 below.

[0165] Table 1

[0166] Method Year ODS OIS AP Human 0.80 0.80 - Canny 1986 0.60 0.63 0.58 Butterfly-shaped 2011 0.59 0.61 0.49 PC / BC 2012 0.64 0.65 0.41 CO 2013 0.64 0.68 0.64 MCI 2014 0.64 0.66 0.56 SCO 2015 0.67 0.71 0.71 SED 2018 0.71 0.74 0.74 MCH 2020 0.70 0.72 0.74 DIDY 2021 0.65 0.66 0.59 This method 0.73 0.75 0.76

[0167] On the NYUDv2 dataset, mean precision and F-score were selected as two common evaluation methods and compared with methods such as CO, MCI, SCO, SED, MCH, and DIDY. The final comparison results are shown in Table 2 below.

[0168] Table 2

[0169]

[0170]

[0171] As can be seen from the two tables, the overall edge detection performance of this invention is significantly better than other methods, both on the early-collected BSDS500 dataset and on the widely used NYUDv2 dataset.

[0172] In this embodiment, a fundamental problem in biomimetic edge detection is how to suppress texture edges while preserving salient region boundaries. This invention utilizes the hierarchical processing of the human visual nervous system to detect edges, including the retina and lateral geniculate body, primary visual cortex, secondary visual cortex, and quaternary visual cortex. To suppress texture edges, a novel peripheral modulation mechanism is designed, including unidirectional facilitation, lateral inhibition, and omnidirectional inhibition. Simultaneously, texture information is used to calculate texture gradients to suppress texture edges and highlight the boundaries of textured regions. Based on the research results of this invention, the proposed biomimetic multi-level edge detection framework achieves state-of-the-art results on two benchmark datasets (BSDS500 and NYUDv2), generating more accurate edge detection maps while detecting and segmenting image background regions with indistinct hierarchies, demonstrating the effectiveness of this method.

Claims

1. A visual biomimetic edge detection method based on texture gradient adjustment, characterized in that, Includes the following steps: S1. Obtain the edge detection image and construct the retinal-coded color, brightness, and brightness contrast channels to obtain multi-channel information; S2. Based on the multi-channel information, construct the classical receptive field of simple cells in the primary visual cortex using the first-order gradient of a two-dimensional Gaussian function, and respond to local edges; S3. Based on the response results, the edge response is adjusted using the surrounding adjustment mechanism, and the texture gradient is introduced to adjust the response of simple cells to obtain the edge response of the primary visual cortex. Step S3 includes the following steps: S301. Based on the response results, a model of the classical receptive field of simple cells is introduced to promote unidirectional synergy: in, Indicates mutual promotion. Indicates the edge orientation is Elliptic Gaussian function, Indicates spatial location, and This indicates the shape of the classical receptive field. , and All represent the coefficients in the elliptic Gaussian function. This represents the response of the classical receptive field; S302. Introducing lateral inhibition to model the classical receptive field of simple cells: in, Indicates lateral inhibition. Indicates the edge orientation is Lateral inhibition nuclei; S303. Introducing peripheral inhibition to model the classic receptive field of simple cells: in, Indicates surrounding inhibition. Indicates a circular average kernel; S304. Based on the modeling results of steps S301-S303, obtain the non-classical receptive field response: in, This indicates a nonclassical receptive field response. , and These represent the weights of unidirectional facilitation, lateral inhibition, and peripheral inhibition, respectively. S305, Introducing a texture kernel: in, Indicates texture kernel, and They represent edge along towards The upper and lower parts, Represents the radius of a local region of a circle; S306. Detect texture region boundaries using multiple texture kernels of different sizes: in, This represents the result of convolving multiple texture kernels of different radius sizes with texture information. This represents the translated texture kernel. This represents the response of the classical receptive field after translation. and express Along the edge towards Directional translation The distance after; S307: By fusing non-classical receptive field responses with texture region boundaries and using nonlinear functions to modulate the responses of simple cells and texture gradients, the edge responses of the primary visual cortex are obtained. in, This represents the response of the primary visual cortex. This indicates the operation of retrieving the maximum value. This represents the result of convolving multiple texture kernels of different radius sizes with texture information. Represents a nonlinear function. Variables representing nonlinear functions, This represents the maximum value of all texture gradients; S4. Based on the edge response of the primary visual cortex, construct endpoint cells and modulate them to obtain the edge response of the secondary visual cortex; S5. Based on the edge response of the second-level visual cortex, the results of all channels are fused to obtain the edge response of the fourth-level visual cortex, thus completing the detection of visual bionic edges.

2. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 1, characterized in that, Step S1 includes the following steps: S101. Obtain the edge detection image and preprocess the edge detection image using the square root function; S102. Based on the preprocessing results, the color information of the edge detection image is encoded using the color antagonism mechanism of the retina; S103. Encode the brightness information and brightness contrast information channels of the edge detection image, and obtain multi-channel information based on the encoding results of step S102.

3. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 2, characterized in that, The preprocessing expression is as follows: in, This represents the red, green, and blue channel information of the edge detection image after preprocessing. This represents the red, green, and blue channel information of the original edge detection image; The expression for encoding the color information of the edge detection image is as follows; in, This indicates a color-antagonistic cell response. Indicates spatial location, Indicates color antagonistic weights. This indicates the preprocessed red channel. This indicates the green channel after preprocessing. Indicates red-green color antagonistic channels. This indicates the blue-yellow color antagonistic channel. This represents the preprocessed blue channel.

4. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 3, characterized in that, The channel response for the brightness information is as follows: The channel response for the brightness and contrast information is as follows: in, Indicates the brightness channel. Indicates the brightness and contrast channel. This represents a circular average kernel. This represents the result of convolution between the brightness channel and a circular average kernel. Represents a circle. Represents the normalization function. This represents the convolution operation. This represents the radius of a local region of a circle.

5. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 1, characterized in that, The response expression of the classical receptive field is as follows: in, This indicates the response of the classical receptive field. To represent partial derivatives, Indicates the direction of the edge. Indicates the direction of the edge towards the total number. Indicates the number of edge orientations. Represents a two-dimensional Gaussian function. This represents the standard deviation of a two-dimensional Gaussian function. This indicates an unbalanced red-green color antagonistic channel. This indicates an unbalanced blue-yellow color antagonistic channel. This indicates a color-antagonistic cell response. Indicates spatial location, and These represent the red-green antagonistic channel and the blue-yellow antagonistic channel, respectively. Indicates the brightness channel. This indicates the brightness / contrast channel.

6. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 1, characterized in that, Step S4 includes the following steps: S401. Based on the edge response of the primary visual cortex, the linear orientation nucleus is divided into left and right parts, and the cells at the left and right endpoints are modeled and denoted as follows: and The expression for the linear orientation towards the kernel is as follows: in, Indicates a linear orientation toward the core. Indicates length, Indicates spatial location, Indicates the direction of the edge; S402, Utilization and Get the response of each location to the edge endpoint: in, Indicates the endpoint response value. Represents the cell nucleus at the left end. Indicates the right-end nucleus. This indicates the response of the primary visual cortex; S403. Based on the acquired endpoint responses, obtain the edge responses of the second-level visual cortex: in, This indicates the response of the second-level visual cortex. Indicates length is The linear orientation towards the core, Indicates the weighting coefficient. Indicates length is The linear orientation towards the core.

7. The visual biomimetic edge detection method based on texture gradient adjustment according to claim 1, characterized in that, The expression for the edge response of the fourth-level visual cortex is as follows: in, This represents the edge response of the fourth level of the visual cortex. This indicates the response of the second-level visual cortex. This indicates a color-antagonistic cell response. Indicates spatial location, and These represent the red-green antagonistic channel and the blue-yellow antagonistic channel, respectively. Indicates the brightness channel. Indicates the brightness and contrast channel. This indicates an unbalanced red-green color antagonistic channel. This indicates an unbalanced blue-yellow color antagonistic channel.

Images