A salient contour perception method based on multivariate connection model of visual computing

Abstract

The present invention discloses a method for saliency contour perception based on a multivariate connection model of visual computing, and constructs a visual computing model with multivariate connection characteristics. In the LGN feedforward connection, the sparse coding characteristics of LGN neurons are simulated, and weight factors are added to achieve preliminary texture suppression and obtain the primary perception result of the contour; in the horizontal connection of the primary visual cortex, the receptive field of the pinwheel-like structure of the primary visual cortex is simulated, and the discharge strength of the central neuron is adjusted based on the distance between neurons and the optimal orientation angle; in the feedback connection of the higher visual cortex, the hue perception characteristics of the higher visual cortex are simulated, and a three-channel hue perception model including surround suppression is constructed to obtain the response of the higher visual cortex to image targets or structures. The present invention constructs a visual computing model with multivariate connection characteristics, so that the acquired contour can effectively highlight the main target while suppressing texture noise.

Description

Technical Field

[0001] The present invention belongs to the field of visual neural computing, and in particular relates to a saliency contour perception method based on a visual computing multivariate connection model. Background Art

[0002] Contour perception is a critical early step in image analysis. Its goal is to extract the target's contour as completely as possible while removing background and texture noise. Traditional contour perception methods primarily rely on differences in pixel color or grayscale values ​​within a local neighborhood. They often overlook the importance of the target's overall information in the local contour detection process. As a result, the resulting contour perception results tend to miss some of the target's structural characteristics and introduce significant texture noise.

[0003] Current contour perception methods that incorporate biological visual mechanisms, such as visual computational models that incorporate mechanisms such as directional selectivity and receptive field coordination, can effectively improve the interpretability and stability of contour perception models. However, these visual computational models focus more on the mechanisms and functions within each functional module in the visual pathway, and usually only consider the feedforward connection characteristics of the visual pathway, simplifying the connection relationship between the functional modules. In fact, during the transmission and processing of visual information in the visual pathway, the primary visual cortex not only receives input from functional modules such as the lateral geniculate nucleus (LGN) in the lower layers of the visual pathway; there are also complex horizontal connections within the primary visual cortex, which is of great significance for improving the adaptability of the response. In addition, the higher-level visual cortex also has feedback connection characteristics to the primary visual cortex, which contributes to the coordinated optimization of the receptive field.

[0004] Therefore, the present invention focuses on the multi-connectivity characteristics of the visual computing model, simulating the connection relationship and interaction of functional modules such as LGN, primary visual cortex and higher visual cortex in the visual pathway, and forming a multi-connectivity relationship including feedforward, horizontal and feedback connections from the perspective of the primary visual cortex. It not only fully reflects the feedforward input information from LGN, but also reflects the dynamic horizontal connection process of neurons in the primary visual cortex. In addition, it also introduces the feedback guidance role of the higher visual cortex for contour perception. Summary of the Invention

[0005] In view of the deficiencies of the existing technology, the present invention proposes a saliency contour perception method based on a multivariate connection model of visual computing.

[0006] The present invention provides a saliency contour perception method based on a visual computing multivariate connection model, the method specifically comprising the following steps:

[0007] Step 1: On the LGN feedforward connection, a sparse measurement model based on the statistical characteristics of the receptive field is constructed to obtain the sparsity profile after the LGN action.

[0008] Step 2: Construct a lateral adjustment model integrating the pinwheel structure receptive field on the level of the primary visual cortex to obtain the contour after the lateral adjustment of the primary visual cortex.

[0009] Step 2.1 Use the two-dimensional Gabor function to simulate the orientation selection characteristics of the classic receptive field of the primary visual cortex, and traverse K filters towards θ k , select e ij (θ k ) is taken as the contour orientation response;

[0010] Step 2.2: Considering that the primary visual cortex has a pinwheel-like functional architecture, that is, the horizontal modulation within the receptive field of the pinwheel structure is related to the distance between neurons and the angle between the optimal response directions, a lateral modulation model integrating the receptive field of the pinwheel structure is constructed based on the horizontal connections of the primary visual cortex.

[0011] First, define the pinwheel structure receptive field window The half window length is σ, and the subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript HC indicates the simulation of the horizontal connection in the primary visual cortex. The row and column coordinates of the surrounding neurons are (i′, j′), and we define The distance between the center and surrounding neurons is shown in formula (10).

[0012]

[0013] Then define for The direction of the connection between the central and surrounding neurons, for Optimal orientation of surrounding neurons; calculation The optimal response direction angle between the central and surrounding neurons is or When the surrounding neurons will enhance the central neurons, the enhancement coefficient ω E (i′, j′) is as shown in formula (11).

[0014]

[0015] Where, Represents the attenuation rate, which determines how quickly the peripheral effect intensity decays as the orientation difference increases. The default value is 3, the same below.

[0016] when or When the surrounding neurons will inhibit the central neurons, the inhibitory effect coefficient ω I(i′, j′) is as shown in formula (12).

[0017]

[0018] The final integrated receptive field window The total enhancement ΔE received by the central neuron is calculated by taking into account the effects of all surrounding neurons within the central neuron. ij and total inhibitory effect ΔI ij , as shown in formula (13).

[0019]

[0020] Calculating contours after lateral modulation of primary visual cortex As shown in formula (14).

[0021]

[0022] Where δ represents the neuron interaction strength coefficient, which is used to adjust the control The value is between 3 and 5. represents the contour orientation response;

[0023] Step 3: Based on the feedback connection of the higher visual cortex, a three-channel higher visual cortex tone perception model including surround suppression is constructed to obtain the contour after the higher visual cortex tone perception.

[0024] Defining the double-antagonistic receptive field window The corresponding inhibitory and excitatory half-window lengths are σ I and σ E , for r + g - 、b + y - Branch and brightness open channel case, σ I =2σ,σ E =σ; for g + r - 、y + b - Branch and brightness closed channel case, σ I =σ,σ E =2σ; subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript G represents the simulation of the feedback connection on the higher visual cortex. The row and column coordinates of the surrounding neurons are (p,q);

[0025] Step 3.1 simulates the hue perception function of the higher visual cortex and constructs a double-antagonistic receptive field model;

[0026] Calculating the competition coefficient of color antagonistic channel neurons and The superscript rg represents r + g - In branch circuit, the superscript gr indicates g + r - In branch circuit, the superscript by indicates b + y - In branch case, the superscript yb indicates y + b - The same applies to branch roads below.

[0027] Considering the interaction between the classical and non-classical receptive fields in the visual cortex, the Gaussian difference function is incorporated into the dual antagonistic receptive field model. For the peripheral neurons with row and column coordinates (p, q), the Gaussian difference function DoG pq The calculation of is shown in formula (15).

[0028]

[0029] With r + g - For example, the competition coefficient of the branch The calculation of is shown in formula (16).

[0030]

[0031] In the formula, the symbol [a] + Indicates the maximum value of 0 and a; R pq , G pq They represent the red and green component inputs corresponding to the surrounding neurons respectively; A1 represents the attenuation coefficient, which defaults to 1.

[0032] Modify the color component input corresponding to the surrounding neurons in formula (16) to calculate g + r - Branch, b + y - Branch and y + b - Neuron competition coefficient of the branch and

[0033] Step 3.2 Calculate the competition coefficient of the brightness on and off channel neurons and

[0034] The brightness-on channel is responsible for enhancing information with brightness higher than the surrounding area, while the brightness-off channel is responsible for enhancing information with brightness lower than the surrounding area. Taking the calculation as an example, the superscript on represents the brightness open channel situation, as shown in formula (17).

[0035]

[0036] Adjust the double antagonistic receptive field window in formula (15) The corresponding inhibitory and excitatory half-windows σ I , σ E , we get the Gaussian difference function DoG corresponding to the brightness closed channel, and then we can get the neuron competition coefficient through formula (17): The superscript off indicates the brightness closed channel situation.

[0037] Step 3.3 simulates the odd and even channel structure of the visual cortex, and integrates the color antagonistic channel information using the local energy model of multi-channel filtering to obtain simple cell activity and

[0038] With r + g - Taking the branch neuron (i, j) as an example, the two-dimensional Gabor filter shown in formula (18) is used to filter the input information, and the information obtained by the filter is combined with the activity of each channel neuron to obtain the odd component simple cell activity Simple cell activity As shown in formula (19).

[0039]

[0040]

[0041] Where, Represents the phase parameter, odd symmetric filter or π / 2, even symmetric filter or π; Represent the double antagonistic receptive field window The odd and even components of the surrounding neurons; A2 represents the model coefficient, which defaults to 1.

[0042] Modify the corresponding neuron competition coefficient in formula (19) to calculate g + r - 、b + y - and y + b - Odd and even component simple cell activity of branches and

[0043] Step 3.4 Calculate the simple cell activity of the brightness channel by fusing the brightness on and off channels and The superscript L represents the brightness channel, as shown in formula (20).

[0044]

[0045] Step 3.5 uses a two-layer complex cell network to simulate the function of the higher visual cortex, process the input information from simple cells, and obtain the contour after the higher visual cortex's color perception.

[0046] The first layer of complex cell networks is responsible for fusing ten groups of simple cell responses of the color antagonism channel and the brightness channel. Unify the contour features extracted from each path, as shown in formula (21).

[0047]

[0048] The second layer of complex cell network adopts the surround inhibition method to achieve texture inhibition effect through competition between neurons, as shown in formula (22).

[0049]

[0050] In the formula, A3 represents the model coefficient, which is set to 1 by default. Indicates the inhibition constant, which defaults to 0.5.

[0051] Step 4: Construct a contour perception model simulating the multivariate connections of the visual cortex to obtain the final saliency contour E ij .

[0052] The response and transmission process of LGN, primary visual cortex and higher visual cortex to visual stimulus signals are simulated to construct the primary response model of image contour, as shown in formula (23).

[0053]

[0054] Where α, β, and γ represent the feedforward, feedback, and horizontal connection coefficients, respectively, which simulate their effect strength in the biological visual neural circuit and are taken as 0.5, 0.25, and 0.25, respectively. τ represents the membrane potential constant of the neuron.

[0055] Image contour primary response Vt ij The value of adopts the first pulse trigger time coding strategy. In order to better correct the background contour and texture noise, the image contour primary response Vt ij Response to higher-level visual cortex Integrate to get the final saliency contour map E ij , as shown in formula (24).

[0056]

[0057] Where A and B represent the integration coefficients, which are both set to 0.5 by default.

[0058] As a preference, the sparse measurement model based on the statistical characteristics of the receptive field is constructed on the LGN feedforward connection to obtain the sparse profile after the LGN effect. Specifically:

[0059] Construct a two-dimensional neuron array Neurons LGN , Array Neurons LGN The number of rows and columns is the same as the external input image I, and the number of rows and columns is recorded as M and N respectively. The superscript LGN represents the simulation on the LGN feedforward connection, and the same below. LGN The model of a single neuron with row and column coordinates (i, j) is shown in formula (1).

[0060]

[0061] Where τ represents the membrane potential constant of the neuron, and LGN Global parameters in the array Neurons LGN The τ value of all neuron models in is the same and is taken as 70 by default; Represents the Neurons located in the array LGN The neuron membrane potential with row and column coordinates (i, j) in I ij Represents the grayscale value of the pixel with row and column coordinates (i, j) in the external input image I.

[0062] Normalized primary sparse profile response The value is obtained by using the first pulse trigger time coding strategy, as shown in equations (2) to (4).

[0063]

[0064]

[0065]

[0066] Where, v spike Indicates the threshold voltage for neuron discharge, the default value is 16; Represents the Neurons located in the array LGN The first discharge time of the neuron with row and column coordinates (i, j); Indicates traversing i,j and taking The maximum value in the array; Indicates traversing i,j and taking The minimum value in the array.

[0067] In order to improve the processing ability of neurons for contour information and texture noise, the traditional sparse metric sparij Based on the consideration Statistical characteristics in the receptive field window, where the variance term It plays the role of distinguishing the contour from the background, and the mean term μ ij It plays a role in balancing the brightness and darkness of the picture. At the same time, combined with the sparse measurement threshold, a new sparse measurement sparsity is obtained. ij , as shown in formulas (5) to (7). Define the receptive field window Used to represent the array Neurons LGN The area centered on the neuron with row and column coordinates (i, j) in , where σ represents the half-window length of the receptive field window, and the same below.

[0068]

[0069]

[0070]

[0071] Where, express In the receptive field window Histogram of n, The dimension of ‖·‖ p Represents the p-norm; σ defaults to 3, the same below.

[0072] Comprehensive consideration of the normalized primary sparse contour response and sparse metrics ij The sparse contour after LGN is obtained by the joint action of As shown in formula (8).

[0073]

[0074] As a preference, the step 2.1 uses a two-dimensional Gabor function to simulate the orientation selection characteristics of the classical receptive field of the primary visual cortex, and traverses K filters toward θ k , select e ij (θ k ) is taken as the contour orientation response; specifically:

[0075] Contour orientation response The calculation of is shown in formula (9).

[0076]

[0077] Where ε represents the spatial compression ratio, which is used to control the aspect ratio of the filter. Its value ranges from 0.23 to 0.92, and the default value is 0.5. K represents the total number of filter orientations, and the default value is 8; max(e ij (θ k )) represents traversing K filters towards θ k , get e ij (θ k ) is the maximum value of .

[0078] The effects of the present invention compared to the prior art are as follows: the present invention constructs a visual computing model with multi-connectivity characteristics, taking into account the feedforward connection from LGN, the horizontal connection from neurons in the same layer of the primary visual cortex, and the feedback connection from the higher-level visual cortex, thereby strengthening the connection relationship between the functional modules of the visual computing model to improve the contour perception performance. In the LGN feedforward connection, the present invention simulates the sparse coding characteristics of LGN neurons. In order to improve the processing ability of LGN neurons for contour information and texture noise, a weight factor is added to the sparse measurement to further enhance the difference between the contour and the background, balance the brightness and darkness levels of the contour, achieve preliminary texture suppression, and obtain the primary perception result of the contour; in the horizontal connection of the primary visual cortex, the present invention simulates the receptive field of the pinwheel-like structure of the primary visual cortex, calculates the enhancement and inhibition effects of non-central neurons on central neurons based on comprehensive factors such as the distance between neurons and the optimal orientation angle, and further adjusts the discharge strength of the central neurons; in the feedback connection of the higher visual cortex, the present invention simulates the hue perception characteristics of the higher visual cortex, constructs a three-channel hue perception model including surround suppression, and obtains the response of the higher visual cortex to image targets or structures through the fusion of the responses of complex cells to the color antagonism channel and the brightness channel and their surround suppression characteristics, effectively highlighting the target contour while suppressing texture noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 Schematic diagram of the relationship between the center and surrounding neurons of the pinwheel structure receptive field. DETAILED DESCRIPTION

[0080] The present invention provides a saliency contour perception method based on a visual computing multivariate connection model, the method specifically comprising the following steps:

[0081] Step 1: On the LGN feedforward connection, a sparse measurement model based on the statistical characteristics of the receptive field is constructed to obtain the sparsity profile after the LGN action.

[0082] The sparse coding characteristics of LGN in visual information transmission help to effectively compress redundant information, thereby quickly extracting the overall outline of the target. LGN , Array Neurons LGNThe number of rows and columns is the same as the external input image I, and the number of rows and columns is recorded as M and N respectively. The superscript LGN represents the simulation on the LGN feedforward connection, and the same below. LGN The model of a single neuron with row and column coordinates (i, j) is shown in formula (1).

[0083]

[0084] Where τ represents the membrane potential constant of the neuron, and LGN Global parameters in the array Neurons LGN The τ value of all neuron models in is the same and is taken as 70 by default; Represents the Neurons located in the array LGN The neuron membrane potential with row and column coordinates (i, j) in I ij Represents the grayscale value of the pixel with row and column coordinates (i, j) in the external input image I.

[0085] Normalized primary sparse profile response The value is obtained by using the first pulse trigger time coding strategy, as shown in equations (2) to (4).

[0086]

[0087]

[0088]

[0089] Where, v spike Indicates the threshold voltage for neuron discharge, the default value is 16; Represents the Neurons located in the array LGN The first discharge time of the neuron with row and column coordinates (i, j); Indicates traversing i,j and taking The maximum value in the array; Indicates traversing i,j and taking The minimum value in the array.

[0090] In order to improve the processing ability of neurons for contour information and texture noise, the present invention uses the traditional sparse metric spar ij Based on the consideration Statistical characteristics in the receptive field window, where the variance term It plays the role of distinguishing the contour from the background, and the mean term μ ij It plays a role in balancing the brightness and darkness of the picture. At the same time, combined with the sparse measurement threshold, a new sparse measurement sparsity is obtained. ij, as shown in formulas (5) to (7). Define the receptive field window Used to represent the array Neurons LGN The area centered on the neuron with row and column coordinates (i, j) in , where σ represents the half-window length of the receptive field window, and the same below.

[0091]

[0092]

[0093]

[0094] Where, express In the receptive field window Histogram of n, The dimension of ‖·‖ p Represents the p-norm; σ defaults to 3, the same below.

[0095] Comprehensive consideration of the normalized primary sparse contour response and sparse metrics ij The sparse contour after LGN is obtained by the joint action of As shown in formula (8).

[0096]

[0097] Step 2: Construct a lateral adjustment model integrating the pinwheel structure receptive field on the level of the primary visual cortex to obtain the contour after the lateral adjustment of the primary visual cortex.

[0098] Step 2.1 Use the two-dimensional Gabor function to simulate the orientation selection characteristics of the classic receptive field of the primary visual cortex, and traverse K filters towards θ k , e ij (θ k ) is taken as the contour orientation response. The calculation of is shown in formula (9).

[0099]

[0100] Where ε represents the spatial compression ratio, which is used to control the aspect ratio of the filter. Its value ranges from 0.23 to 0.92, and the default value is 0.5. K represents the total number of filter orientations, and the default value is 8; max(e ij (θ k )) represents traversing K filters towards θ k , get e ij (θk ) is the maximum value of .

[0101] Step 2.2 Considering that the primary visual cortex has a pinwheel-like functional structure, that is, the horizontal modulation effect within the receptive field of the pinwheel structure is related to the distance between neurons and the angle between the optimal response directions, a lateral regulation model integrating the receptive field of the pinwheel structure is constructed on the horizontal connection of the primary visual cortex.

[0102] First, define the pinwheel structure receptive field window The half window length is σ, and the subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript HC indicates the simulation of the horizontal connection in the primary visual cortex. The row and column coordinates of the surrounding neurons are (i′, j′), and we define The distance between the center and surrounding neurons is shown in formula (10).

[0103]

[0104] Then define for The direction of the connection between the central and surrounding neurons, for The optimal orientation of surrounding neurons, such as Figure 1 Calculation The optimal response direction angle between the central and surrounding neurons is or When the surrounding neurons will enhance the central neurons, the enhancement coefficient ω E (i′, j′) is as shown in formula (11).

[0105]

[0106] Where, Represents the attenuation rate, which determines how quickly the peripheral effect intensity decays as the orientation difference increases. The default value is 3, the same below.

[0107] when or When the surrounding neurons will inhibit the central neurons, the inhibitory effect coefficient ω I (i′, j′) is as shown in formula (12).

[0108]

[0109] The final integrated receptive field window The total enhancement ΔE received by the central neuron is calculated by taking into account the effects of all surrounding neurons within the central neuron. ij and total inhibitory effect ΔI ij, as shown in formula (13).

[0110]

[0111] Calculating contours after lateral modulation of primary visual cortex As shown in formula (14).

[0112]

[0113] Where δ represents the neuron interaction strength coefficient, which is used to adjust the control The value is between 3 and 5.

[0114] Step 3: Based on the feedback connection of the higher visual cortex, a three-channel higher visual cortex tone perception model including surround suppression is constructed to obtain the contour after the higher visual cortex tone perception.

[0115] Defining the double-antagonistic receptive field window The corresponding inhibitory and excitatory half-window lengths are σ I and σ E , for r + g - 、b + y - Branch and brightness open channel case, σ I =2σ,σ E =σ; for g + r - 、y + b - Branch and brightness closed channel case, σ I =σ,σ E =2σ; subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript G represents the simulation of the feedback connection on the higher visual cortex. The row and column coordinates of the surrounding neurons are (p,q).

[0116] Step 3.1: Simulate the hue perception function of the higher visual cortex and construct a dual-antagonistic receptive field model. Calculate the competition coefficient of the color antagonistic channel neurons. and The superscript rg represents r + g - In branch circuit, the superscript gr indicates g + r - In branch circuit, the superscript by indicates b + y - In branch case, the superscript yb indicates y + b - The same applies to branch roads below.

[0117] Considering the interaction between the classical and non-classical receptive fields in the visual cortex, the Gaussian difference function is incorporated into the dual antagonistic receptive field model. For the peripheral neurons with row and column coordinates (p, q), the Gaussian difference function DoG pq The calculation of is shown in formula (15).

[0118]

[0119] With r + g - For example, the competition coefficient of the branch The calculation of is shown in formula (16).

[0120]

[0121] In the formula, the symbol [a] + Indicates the maximum value of 0 and a; R pq , G pq They represent the red and green component inputs corresponding to the surrounding neurons respectively; A1 represents the attenuation coefficient, which defaults to 1.

[0122] Modify the color component input corresponding to the surrounding neurons in formula (16) to calculate g + r - Branch, b + y - Branch and y + b - Neuron competition coefficient of the branch and

[0123] Step 3.2 Calculate the competition coefficient of the brightness on and off channel neurons and The brightness-on channel is responsible for enhancing information with brightness higher than the surrounding area, while the brightness-off channel is responsible for enhancing information with brightness lower than the surrounding area. Taking the calculation as an example, the superscript on represents the brightness open channel situation, as shown in formula (17).

[0124]

[0125] Adjust the double antagonistic receptive field window in formula (15) The corresponding inhibitory and excitatory half-windows σ I , σ E , we get the Gaussian difference function DoG corresponding to the brightness closed channel, and then we can get the neuron competition coefficient through formula (17): The superscript off indicates the brightness closed channel situation.

[0126] Step 3.3 simulates the odd and even channel structure of the visual cortex, and integrates the color antagonistic channel information using the local energy model of multi-channel filtering to obtain simple cell activity and

[0127] With r + g - Taking the branch neuron (i, j) as an example, the two-dimensional Gabor filter shown in formula (18) is used to filter the input information, and the information obtained by the filter is combined with the activity of each channel neuron to obtain the odd component simple cell activity Simple cell activity As shown in formula (19).

[0128]

[0129]

[0130] Where, Represents the phase parameter, odd symmetric filter or π / 2, even symmetric filter or π; Represent the double antagonistic receptive field window The odd and even components of the surrounding neurons; A2 represents the model coefficient, which defaults to 1.

[0131] Modify the corresponding neuron competition coefficient in formula (19) to calculate g + r - 、b + y - and y + b - Odd and even component simple cell activity of branches and

[0132] Step 3.4 Calculate the simple cell activity of the brightness channel by fusing the brightness on and off channels and The superscript L represents the brightness channel, as shown in formula (20).

[0133]

[0134] Step 3.5 uses a two-layer complex cell network to simulate the function of the higher-level visual cortex, process the input information from simple cells, and obtain the higher-level visual cortex response

[0135] The first layer of complex cell network is responsible for fusing the ten groups of simple cell responses of the color antagonism channel and the brightness channel, and unifying the contour features extracted from each channel, as shown in Equation (21).

[0136]

[0137] The second layer of complex cell network adopts the surround inhibition method to achieve texture inhibition effect through competition between neurons, as shown in formula (22).

[0138]

[0139] In the formula, A3 represents the model coefficient, which is set to 1 by default. Indicates the inhibition constant, which defaults to 0.5.

[0140] Step 4: Construct a contour perception model simulating the multivariate connections of the visual cortex to obtain the final saliency contour E ij .

[0141] The response and transmission process of LGN, primary visual cortex and higher visual cortex to visual stimulus signals are simulated to construct the primary response model of image contour, as shown in formula (23).

[0142]

[0143] Where α, β, and γ represent the feedforward, feedback, and horizontal connection coefficients, respectively, which simulate their effect strength in the biological visual neural circuit and are taken as 0.5, 0.25, and 0.25, respectively.

[0144] Image contour primary response Vt ij The value of adopts the first pulse trigger time coding strategy shown in formulas (2) to (4), V ij (t) Analogy to Vt ij Analogy In order to better correct the background contour and texture noise, the image contour primary response Vt ij Response to higher-level visual cortex Integrate to get the final saliency contour map E ij , as shown in formula (24).

[0145]

[0146] Where A and B represent the integration coefficients, which are both set to 0.5 by default.

Claims

1. A salient contour perception method based on a multivariate connection model of visual computing, characterized in that: The method specifically comprises the following steps: Step 1: On the LGN feedforward connection, a sparse measurement model based on the statistical characteristics of the receptive field is constructed to obtain the sparsity profile after the LGN action. Specifically: Construct a two-dimensional neuron array Neurons LGN , Array Neurons LGN The number of rows and columns is the same as the external input image I, and the number of rows and columns is recorded as M, N respectively. The superscript LGN represents the simulation on the LGN feedforward connection, and the same below. The array Neurons LGN The single neuron model with row and column coordinates (i, j) is shown in formula (1); Where τ represents the membrane potential constant of the neuron, and LGN Global parameters in the array Neurons LGN The τ value of all neuron models in is the same; Represents the Neurons located in the array LGN The neuron membrane potential with row and column coordinates (i, j) in I ij Represents the grayscale value of the pixel with row and column coordinates (i, j) in the external input image I; Normalized primary sparse profile response The value is obtained by using the first pulse trigger time coding strategy, as shown in equations (2) to (4); Where, v spike represents the threshold voltage for neuronal firing; Represents the Neurons located in the array LGN The first discharge time of the neuron with row and column coordinates (i, j); Indicates traversing i,j and taking The maximum value in the array; Indicates traversing i,j and taking The minimum value in the array; In order to improve the processing ability of neurons for contour information and texture noise, the traditional sparse metric spar ij Based on the consideration Statistical characteristics in the receptive field window, where the variance term It plays the role of distinguishing the contour from the background, and the mean term μ ij It plays a role in balancing the brightness and darkness of the picture; at the same time, it combines the sparse measurement threshold to obtain a new sparse measurement sparsity ij , as shown in formulas (5) to (7); define the receptive field window Used to represent the array Neurons LGN The area centered on the neuron with row and column coordinates (i, j) is σ, where σ represents the half-window length of the receptive field window, the same below; Where, express In the receptive field window Histogram of n, The dimension of ‖·‖ p represents the p-norm; Comprehensive consideration of the normalized primary sparse contour response and sparse metrics ij The sparse contour after LGN is obtained by the joint action of As shown in formula (8); Step 2: Construct a lateral adjustment model integrating the pinwheel structure receptive field on the level of the primary visual cortex to obtain the contour after the lateral adjustment of the primary visual cortex. Step 2.1 Use the two-dimensional Gabor function to simulate the orientation selection characteristics of the classic receptive field of the primary visual cortex, and traverse K filters towards θ k , select e ij (θ k ) is taken as the contour orientation response; specifically: Contour orientation response The calculation of is shown in formula (9); Where ε represents the spatial compression ratio, which is used to control the aspect ratio of the filter and has a value in the range of 0.23 to 0.

92. K represents the total number of filter orientations; max(e ij (θ k )) represents traversing K filters towards θ k , get e ij (θ k )'s maximum value; Step 2.2: Considering that the primary visual cortex has a pinwheel-like functional architecture, that is, the horizontal modulation within the receptive field of the pinwheel structure is related to the distance between neurons and the angle between the optimal response directions, a lateral modulation model integrating the receptive field of the pinwheel structure is constructed based on the horizontal connections of the primary visual cortex. First, define the pinwheel structure receptive field window The half window length is σ, and the subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript HC indicates the simulation at the level of the primary visual cortex connection; The row and column coordinates of the surrounding neurons are (i′, j′), and we define The distance between the center and surrounding neurons is shown in formula (10); Then define for The direction of the connection between the central and surrounding neurons, for Optimal orientation of surrounding neurons; calculation The optimal response direction angle between the central and surrounding neurons is or When the surrounding neurons will enhance the central neurons, the enhancement coefficient ω E (i′, j′) is as shown in formula (11); Where, represents the decay rate; when or When the surrounding neurons will inhibit the central neurons, the inhibitory effect coefficient ω I (i′, j′) is as shown in formula (12); The final integrated receptive field window The total enhancement ΔE received by the central neuron is calculated by taking into account the effects of all surrounding neurons within the central neuron. ij and total inhibitory effect ΔI ij , as shown in formula (13); Calculating contours after lateral modulation of primary visual cortex As shown in formula (14); Where δ represents the neuron interaction strength coefficient, which is used to adjust the control The value is between 3 and 5. represents the contour orientation response; Step 3: Based on the feedback connection of the higher visual cortex, a three-channel higher visual cortex tone perception model including surround suppression is constructed to obtain the contour after the higher visual cortex tone perception. Defining the double-antagonistic receptive field window The corresponding inhibitory and excitatory half-window lengths are σ I and σ E , for r + g - 、b + y - Branch and brightness open channel case, σ I =2σ,σ E =σ; for g + r - 、y + b - Branch and brightness closed channel case, σ I =σ,σ E =2σ; subscript ij represents The row and column coordinates of the central neuron are (i, j), and the superscript G represents the simulation of the feedback connection in the higher visual cortex; assuming The row and column coordinates of the surrounding neurons are (p,q); Step 3.1 simulates the hue perception function of the higher visual cortex and constructs a double-antagonistic receptive field model; Calculating the competition coefficient of color antagonistic channel neurons and The superscript rg represents r + g - In branch circuit, the superscript gr indicates g + r - In branch circuit, the superscript by indicates b + y - In branch case, the superscript yb indicates y + b - The same applies to branch roads below; Considering the interaction between the classical and non-classical receptive fields in the visual cortex, the Gaussian difference function is incorporated into the double antagonistic receptive field model; for the peripheral neurons with row and column coordinates (p, q), the Gaussian difference function DoG pq The calculation of is shown in formula (15); With r + g - For example, the competition coefficient of the branch The calculation of is shown in formula (16); In the formula, the symbol [a] + Indicates the maximum value of 0 and a; R pq , G pq They represent the red and green component inputs corresponding to the surrounding neurons respectively; A1 represents the attenuation coefficient; Modify the color component input corresponding to the surrounding neurons in formula (16) to calculate g + r - Branch, b + y - Branch and y + b - Neuron competition coefficient of the branch and Step 3.2 Calculate the competition coefficient of the brightness on and off channel neurons and The brightness-on channel is responsible for enhancing the information of the area with higher brightness than the surrounding area, while the brightness-off channel is responsible for enhancing the information of the area with lower brightness than the surrounding area. Taking the calculation as an example, the superscript on represents the brightness open channel situation, as shown in formula (17); Adjust the double antagonistic receptive field window in formula (15) The corresponding inhibitory and excitatory half-windows σ I , σ E , we get the Gaussian difference function DoG corresponding to the brightness closed channel, and then we can get the neuron competition coefficient through formula (17): The superscript off indicates the brightness closed channel situation; Step 3.3 simulates the odd and even channel structure of the visual cortex, and integrates the color antagonistic channel information using the local energy model of multi-channel filtering to obtain simple cell activity and For r + g - The branch neuron (i, j) uses the two-dimensional Gabor filter shown in formula (18) to filter the input information, and combines the information obtained by the filter with the activity of each channel neuron to obtain the odd component simple cell activity Simple cell activity As shown in formula (19); Where, Represents the phase parameter, odd symmetric filter or π / 2, even symmetric filter or π; Represent the double antagonistic receptive field window The odd and even components of the surrounding neurons; A2 represents the model coefficient; Modify the corresponding neuron competition coefficient in formula (19) to calculate g + r - 、b + y - and y + b - Odd and even component simple cell activity of branches and Step 3.4 Calculate the simple cell activity of the brightness channel by fusing the brightness on and off channels and The superscript L represents the brightness channel case, as shown in formula (20); Step 3.5 uses a two-layer complex cell network to simulate the function of the higher visual cortex, process the input information from simple cells, and obtain the contour after the higher visual cortex's color perception. The first layer of complex cell networks is responsible for fusing ten groups of simple cell responses of the color antagonism channel and the brightness channel. Unify the contour features extracted from each path, as shown in formula (21); The second layer of complex cell network adopts the surround inhibition method to achieve texture inhibition effect through competition between neurons, as shown in formula (22); Where A3 represents the model coefficient, represents the inhibition constant; Step 4: Construct a contour perception model simulating the multivariate connections of the visual cortex to obtain the final saliency contour E ij ; Simulate the response and transmission process of LGN, primary visual cortex and higher visual cortex to visual stimulus signals, and construct the primary response model of image contour, as shown in formula (23); Where α, β, and γ represent the feedforward, feedback, and horizontal connection coefficients, respectively, simulating their effect strength in the biological visual neural circuit, and τ represents the membrane potential constant of the neuron; Image contour primary response Vt ij The value of adopts the first pulse trigger time coding strategy. In order to better correct the background contour and texture noise, the image contour primary response Vt ij Response to higher-level visual cortex Integrate to get the final saliency contour map E ij , as shown in formula (24); Where A and B represent the integration coefficients respectively.

2. The method for salient contour perception based on a multivariate connectivity model of visual computing according to claim 1, characterized in that: The is 3.

3. The method for salient contour perception based on a multivariate connectivity model of visual computing according to claim 1, characterized in that: A1, A2, A3 take 1, Take 0.5, α, β and γ as 0.5, 0.25 and 0.25 respectively, and A and B are set to 0.

5.

4. The method for salient contour perception based on a multivariate connectivity model of visual computing according to claim 1, characterized in that: τ is taken as 70; v spike Take 16 and σ take 3.

5. The method for salient contour perception based on a multivariate connectivity model of visual computing according to claim 1, characterized in that: ε is 0.5 and K is 8.

Images