CN104599262A - Multichannel pulse coupling neural network based color image segmentation technology - Google Patents

Abstract

A multichannel pulse coupling neural network based color image segmentation technology comprises the steps of step (1), inputting images to be segmented; step (2), using color vectors of all pixels of the images as input vectors of one input neuron and eight adjacent pixel color vectors as radial basis function (RBF) characteristic vectors, and determining initial seed points through seed selection conditions; step (3), growing the seed region through growing rules, classifying the pixel points in accordance with the growing rules in the seed region, and connecting the neurons and grouping and numbering the neurons; step (4), calculating the average characteristic vector of all connection regions, and replacing the characteristic vectors included in all neurons of the region with the obtained characteristic vector; step (5), connecting the qualified unconnected neurons with the proximate groups through a rapid connection rule; step (6),updating the preset threshold to be theta i = theta i1 +delta theta i, and repeating the step (5); step (7), performing rule merging on accordant regions in the images and merging proximate region blocks on space; repeating the step (7) till the region merging stopping conditions are met to complete color image segmentation.

Description

Based on the color Image Segmentation of hyperchannel Pulse Coupled Neural Network

Technical field

The present invention relates to a kind of color Image Segmentation for digital image processing field, specifically a kind of parallel computing of the color images based on hyperchannel Pulse Coupled Neural Network.

Background technology

Iamge Segmentation is the technology by display foreground and background segment, utilizes Iamge Segmentation often can extract foreground target or area-of-interest further from image.Through the development of decades, image Segmentation Technology is divided into four classes: threshold method, the method based on edge, the method based on region and mixed method.In numerous image partition methods, Pulse Coupled Neural Network (PCNN) is a kind of biological heuristic.The beginning of the nineties, Germany scientist professor and his co-worker find when zoologizeing optic nerve model, if the nervous centralis sensing region in certain limit perceives the signal excitation of feature similarity, this excitation just can cause synchronized oscillation in that region, follow-up research shows, this resonance model that causes is applicable to Iamge Segmentation.

Traditional PCNN model is confined in gray level image process, its limiting factors mainly contains 2 points: first, the essence of intrerneuron coupling is confined to the Scalar operation of grey scale pixel value, PCNN is made to be unsuitable for solving coloured image problem, this is because color must with vector representation (such as red green blue tricolor combination); The second, algorithm complex is excessive.

For these limitations, the present invention creates a kind of new pulse god coupling through network model, and the parallel neural network algorithm developed based on programmable gate array (FPGA) and neuron circuit, this platform realizes the color images of quick high-quality.

Summary of the invention

The present invention will overcome the problem existing for conventional P CNN, be intended to accelerate algorithm speed and ensure segmentation quality, proposing a kind of based on hyperchannel Pulse Coupled Neural Network (English full name: multi-channel pulse coupled neural network; English abbreviation: MPCNN) color Image Segmentation.

First new method utilizes MPCNN that color image pixel information is inputted neuron arrays one to one, then all neurons are operated with running simultaneously, selected seed neuron, connection dividing into groups, finally carry out region growing and merge obtaining segmentation result.The topological structure of hyperchannel Pulse Coupled Neural Network as shown in Figure 1.The method selects RBF (radial basis function) neuron to solve the calculating of color vector distance between pixel, and in the upper realization of FPGA (can gate array be edited).Parallel synchronous operation can be carried out to all pixels, utilize pixel color proper vector to realize pulse-couple.Under the prerequisite ensureing segmentation quality, greatly reduce the processing time.The method, mainly for the segmentation of coloured image, all has good segmentation effect for most of coloured image, for successive image work for the treatment of provides reliable basis.

Above-mentioned RBF neuron models (see Fig. 2) are described below:

Input neuron u _i(i=1 ... N _i) by cynapse and output neuron v _j(j=1 ... N _j) be closely connected, wherein i, j are respectively input, output neuron subscript, N _i, N _jbe respectively input, output neuron number.For any one cynapse, when a postsynaptic potential (PSP) is excited, the distance between input neuron and cynapse value and 8-neighborhood territory pixel determines the start time.

Wherein PSP _i ^jt () is postsynaptic potential, j be neighbor subscript (1 ..., 8), t is the time, x _ifor pixel color vector, represent RGB (RGB) information, for neuronic proper vector.

When in the time period time interior, postsynaptic potential and be

$s_j\left(t\right)=\underset{i=1}{\overset{N_i}{\mathrm{\Σ}}}{{\mathrm{PSP}}_i}^j\left(t\right)=\underset{i=1}{\overset{N_i}{\mathrm{\Σ}}}(t-|x_i-c_i^j\left|\right)---\left(2\right)$

Wherein s _jt cumulative sum that () is postsynaptic potential.

After all postsynaptic potentials are all excited, export pulses generation.Suppose to work as s _jwhen () reaches a threshold value θ t, output neuron is at t _jmoment produces pulse,

$t_j=\frac{\mathrm{\θ}}{N_i}+\frac{1}{N_i}\underset{i=1}{\overset{N_i}{\mathrm{\Σ}}}|x_i-c_i^j|---\left(3\right)$

Wherein t _jfor the time, θ is threshold value.T _jreflect one about x and c ^jmanhatton distance radial basis function,

$\left|\right|x-c^j\left|\right|=\underset{i=1}{\overset{N_i}{\mathrm{\Σ}}}|x_i-c_i^j|---\left(4\right)$

Wherein || x-c ^j|| be manhatton distance.

The Segmentation of Color Image of above-mentioned hyperchannel Pulse Coupled Neural Network, step is as follows:

(1) image to be split is inputted;

(2) using the input vector of the color-vector of each pixel of image as an input neuron, 8 neighbor color-vectors, as each proper vector of RBF, utilize initial point selection condition determination initial seed point;

(3) utilize growing strategy to grow seed region, the pixel meeting growing strategy is included in seed region, namely neuron is connected, and packet numbering;

(4) calculate the averaged feature vector of each join domain, and replace the proper vector comprised in all neurons in this region by the proper vector obtained;

(5) if there is the neuron do not connected, step (6) is performed, otherwise, perform step (7);

(6) utilize quick concatenate rule, connect simultaneously and qualified do not connect neuron to immediate grouping; Predetermined threshold is updated to θ _i=θ _il+ Δ θ _i, wherein θ _ilfor former threshold value, Δ θ _ifor increment, θ _ifor upgrading rear threshold value, repeat step (5);

(7) utilize region merging technique rule, to meeting region merging technique rule in image, and spatially close region unit merges;

(8) step (7) is repeated, until meet region merging technique stop condition;

(9) color images is completed.

In step (2), choose in process at Seed Points, sub pixel must keep similarity with its neighborhood territory pixel to a certain extent.In RGB color space, if the ultimate range of a pixel and its 8-neighborhood territory pixel point is less than predetermined threshold value, then this pixel can as a Seed Points, and corresponding neuron is exactly Seed Points neuron.For arbitrary pixel s, if meet formula (5), then can be used as Seed Points:

μ _s<θ _μ；

${\mathrm{\&mu;}}_s=\underset{j=1,...,8}{\max }\left\{\right||x^s-c^j|\left|\right\}---\left(5\right)$

Wherein, s is pixel label, and j is neighbor subscript, || x ^s-c ^j|| be manhatton distance, θ _μfor predetermined threshold value, μ _sfor the ultimate range of pixel s and 8-neighborhood territory pixel point.

In step (3), in the process of region growing, be as starting point using selected Seed Points, (5) are as judgment condition with the formula, to the growth of its 8-neighborhood, until pixel is not meeting formula (5), by the group areas numbering obtained.

In step (4), calculate the averaged feature vector of each grouping according to formula (6),

${\mathrm{\&sigma;}}^g=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{x}_j^g---\left(6\right)$

${\mathrm{\&sigma;}}^g={[{\mathrm{\&sigma;}}_R^g,{\mathrm{\&sigma;}}_G^g,{\mathrm{\&sigma;}}_B^g]}^T$

Wherein σ ^gfor the mean eigenvalue of RGB color space, be respectively RGB component mean value, M is the pixel number of the grouping being numbered g, and g is the numbering of grouping, and j is neighbor subscript.

In step (6), for arbitrary neuron u do not connected, be connected in immediate neighboring region according to formula (7),

$j_u=\underset{j}{\arg }\left(\underset{j=1,...,8}{\min }\right||x^u-{\mathrm{\&sigma;}}^{g_j})---\left(7\right)$

Wherein, j _ufor connecting pixel subscript, x ^uthe proper vector of this pixel, that neighboring region is numbered g _javeraged feature vector.All neurons that do not connect are operated, and the difference not connecting neuron and neighboring region must be less than defined threshold simultaneously, and upgrade threshold value:

θ _i＝θ _i+Δθ _i(8)

Wherein, θ _ifor defined threshold, Δ θ _ifor threshold delta.

In step (7), in region merging technique process, using region distance and area size as two judgment condition:

(1) if the manhatton distance of two adjacent areas is less than predetermined threshold value, be a region by two region merging technique, and zoning average again;

(2) if area size is less than predetermined threshold value, this region is incorporated to the adjacent area minimum with its manhatton distance.

The process flow diagram of above-mentioned algorithm as shown in Figure 3.

Advantage of the present invention is: accelerate algorithm speed and ensure Iamge Segmentation quality.

Accompanying drawing explanation

Fig. 1 is the topological structure structural drawing of hyperchannel Pulse Coupled Neural Network of the present invention

Fig. 2 is pulse RBF neural model.

Fig. 3 is the algorithm flow chart adopting the inventive method.

Fig. 4 is original color image.

Fig. 5 is the result of Iamge Segmentation.

Embodiment

The enforcement item of color image segmentation method is illustrated below in conjunction with accompanying drawing.

The invention process is in the system based on Spatan-3A FPGA (XC3SD3400A-4FGG676C) platform, and this FPGA has 3,400,000 logic gates, and working clock frequency is 250MHz.In implementation process, cutting procedure has used the FPGA resource of about 75 %.

(1) image to be split is inputted, as shown in Figure 4;

(2) using the proper vector of the color-vector of each pixel of image as an input neuron, utilize initial point selection condition determination initial seed point, each neuron circuit can hardware run parallel synchronous;

(3) utilize growing strategy to grow seed region, the pixel ultimate range with its 8-neighborhood territory pixel point being less than predetermined threshold value can be included in seed region as a Seed Points, namely connects neuron, and packet numbering.Can synchronous operation on hardware;

(4) calculate the averaged feature vector of each join domain, and replace the proper vector comprised in all neurons in this region by the proper vector obtained;

(5) if there is the neuron do not connected, step (6) is performed, otherwise, perform step (7);

(6) utilize quick concatenate rule, connect simultaneously and qualified do not connect neuron to immediate grouping; Predetermined threshold is updated to θ _i=θ _il+ Δ θ _i, wherein θ _ilfor former threshold value, Δ θ _ifor increment, θ _ifor upgrading rear threshold value, repeat step (5);

(7) utilize region merging technique rule, to meeting region merging technique rule in image, and spatially close region unit merges;

(8) step (7) is repeated, until meet region merging technique stop condition to choose how group threshold value repeatedly merges image;

(9) color images is completed.

In Figure 5, result shows this method has better segmentation result to coloured image in segmentation result display.Algorithm is implemented on FPGA, to accelerate sliced time, for follow-up real-time process provides Reliable guarantee.

Above-mentioned RBF neuron models (see Fig. 2) are described below:

Input neuron u _i(i=1 ... N _i) by cynapse and output neuron v _j(j=1 ... N _j) be closely connected.For any one cynapse, when a postsynaptic potential (PSP) is excited, the distance between input neuron and cynapse value and 8-neighborhood territory pixel determines the start time.

Wherein for postsynaptic potential, j be neighbor subscript (1 ..., 8), t is the time, x _ifor pixel color vector, represent RGB information, for neuronic proper vector.

When in the time period time interior, postsynaptic potential and be

$s_j\left(t\right)=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{{\mathrm{PSP}}_i}^j=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}(t-|x_i-c_i^j\left|\right)---\left(2\right)$

Wherein s _jt cumulative sum that () is postsynaptic potential.

After all postsynaptic potentials are all excited, export pulses generation.Suppose to work as s _jwhen () reaches a threshold value θ t, output neuron is at t _jmoment produces pulse,

$t_j=\frac{\mathrm{\&theta;}}{N_i}+\frac{1}{N_i}\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}|x_i-c_i^j|---\left(3\right)$

Wherein t _jfor the time, θ is threshold value.T _jreflect one about x and c ^jmanhatton distance radial basis function,

$\left|\right|x-c^j\left|\right|=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}|x_i-c_i^j|---\left(4\right)$

Wherein || x-c ^j|| be manhatton distance.

In step (2), choose in process at Seed Points, sub pixel must keep similarity with its neighborhood territory pixel to a certain extent.In RGB color space, if the ultimate range of a pixel and its 8-neighborhood territory pixel point is less than predetermined threshold value, then this pixel can as a Seed Points, and corresponding neuron is exactly Seed Points neuron.For arbitrary pixel s, if meet formula (5), then can be used as Seed Points:

$\begin{array}{c}{\mathrm{\&mu;}}_s<{\mathrm{\&theta;}}_{\mathrm{\&mu;}}\\ {\mathrm{\&mu;}}_s=\underset{j=1,...,8}{\max }\left\{\right||x^s-c^j|\left|\right\}\end{array}---\left(5\right)$

Wherein, s is pixel label, and j is neighbor subscript, || x ^s-c ^j|| be manhatton distance, θ _μfor predetermined threshold value, μ _sfor the ultimate range of pixel s and 8-neighborhood territory pixel point.

In step (3), in the process of region growing, be as starting point using selected Seed Points, (5) are as judgment condition with the formula, to the growth of its 8-neighborhood, until pixel is not meeting formula (5), by the group areas numbering obtained.

In step (4), calculate the averaged feature vector of each grouping according to formula (6),

${\mathrm{\&sigma;}}^g=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{x}_j^g---\left(6\right)$

${\mathrm{\&sigma;}}^g={[{\mathrm{\&sigma;}}_R^g,{\mathrm{\&sigma;}}_G^g,{\mathrm{\&sigma;}}_B^g]}^T$

Wherein σ ^gfor the mean eigenvalue of RGB color space, be respectively RGB component mean value, M is the pixel number of the grouping being numbered g, and g is the numbering of grouping, and j is neighbor subscript.

In step (6), for arbitrary neuron u do not connected, be connected in immediate neighboring region according to formula (7),

$j_u=\underset{j}{\arg }\left(\underset{j=1,...,8}{\min }\right||x^u-{\mathrm{\&sigma;}}^{g_j}|\left|\right)---\left(7\right)$

Wherein, u is for connecting pixel, x ^uthe proper vector of this pixel, that neighboring region is numbered g _javeraged feature vector.All neurons that do not connect are operated, and the difference not connecting neuron and neighboring region must be less than defined threshold simultaneously, and upgrade threshold value:

θ _i＝θ _i+Δθ _i(8)

Wherein, θ _ifor defined threshold, Δ θ _ifor threshold delta.

In step (7), in region merging technique process, using region distance and area size as two judgment condition:

(1) if the manhatton distance of two adjacent areas is less than predetermined threshold value, be a region by two region merging technique, and zoning average again;

(2) if area size is less than predetermined threshold value, this region is incorporated to the adjacent area minimum with its manhatton distance.

The process flow diagram of above-mentioned algorithm as shown in Figure 3.

Content described in this instructions embodiment is only enumerating the way of realization of inventive concept; protection scope of the present invention should not be regarded as being only limitted to the concrete form that embodiment is stated, protection scope of the present invention also and conceive the equivalent technologies means that can expect according to the present invention in those skilled in the art.

Claims (6)

1. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network, step is as follows:

(1) image to be split is inputted;

(2) using the input vector of the color-vector of each pixel of image as an input neuron, 8 neighbor color-vectors, as each proper vector of radial basis function RBF, utilize initial point selection condition determination initial seed point;

(3) utilize growing strategy to grow seed region, the pixel meeting growing strategy is included in seed region, namely neuron is connected, and packet numbering;

(4) calculate the averaged feature vector of each join domain, and replace the proper vector comprised in all neurons in this region by the proper vector obtained;

(5) if there is the neuron do not connected, step (6) is performed, otherwise, perform step (7);

(6) utilize quick concatenate rule, connect simultaneously and qualified do not connect neuron to immediate grouping; Predetermined threshold is updated to θ _i=θ _il+ Δ θ _i, wherein θ _ilfor former threshold value, Δ θ _ifor increment, θ _ifor upgrading rear threshold value, repeat step (5);

(7) utilize region merging technique rule, to meeting region merging technique rule in image, and spatially close region unit merges;

(8) step (7) is repeated, until meet region merging technique stop condition;

(9) color images is completed.

2. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network as claimed in claim 1, it is characterized in that: in step (2), choose in process at Seed Points, sub pixel must keep similarity with its neighborhood territory pixel to a certain extent; In RGB RGB color space, if the ultimate range of a pixel and its 8-neighborhood territory pixel point is less than predetermined threshold value, then this pixel can as a Seed Points, and corresponding neuron is exactly Seed Points neuron; For arbitrary pixel s, if meet formula (5), then can be used as Seed Points:

μ _s<θ _μ

${\mathrm{\&mu;}}_s=\underset{j=1,...,8}{\max }\left\{\right||x^s-c^j|\left|\right\}---\left(5\right)$

Wherein, s is pixel label, and j is neighbor subscript, ‖ x ^s-c ^j‖ is manhatton distance, θ _μfor predetermined threshold value, μ _sfor the ultimate range of pixel s and 8-neighborhood territory pixel point;

Described RBF neuron models are described below:

Input neuron u _i(i=1 ... N _i) by cynapse and output neuron v _j(j=1 ... N _j) be closely connected, wherein i, j are respectively input, output neuron subscript, N _i, N _jbe respectively input, output neuron number; For any one cynapse, when a postsynaptic potential (PSP) is excited, the distance between input neuron and cynapse value and 8-neighborhood territory pixel determines the start time;

Wherein PSP _i ^jt () is postsynaptic potential, j be neighbor subscript (1 ..., 8), t is the time, x _ifor pixel color vector, represent RGB information, for neuronic proper vector;

When in the time period time interior, postsynaptic potential and be:

$s_j\left(t\right)=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{{\mathrm{PSP}}_i}^j\left(t\right)=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}(t-|x_i-c_i^j\left|\right)---\left(2\right)$

Wherein s _jt cumulative sum that () is postsynaptic potential;

After all postsynaptic potentials are all excited, export pulses generation; Suppose to work as s _jwhen () reaches a threshold value θ t, output neuron is at t _jmoment produces pulse,

$t_j=\frac{\mathrm{\&theta;}}{N_i}+\frac{1}{N_i}\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}|x_i-c_i^j|---\left(3\right)$

Wherein t _jfor the time, θ is threshold value; t _jreflect one about x and c ^jmanhatton distance radial basis function,

$\left|\right|x-c^j\left|\right|=\underset{i=1}{\overset{N_i}{\mathrm{\&Sigma;}}}|x_i-c_i^j|---\left(4\right)$

Wherein ‖ x-c ^j‖ is manhatton distance.

3. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network as claimed in claim 1, it is characterized in that: in step (3), in the process of region growing, as starting point using selected Seed Points, (5) are as judgment condition with the formula, to the growth of its 8-neighborhood, until pixel is not meeting formula (5), by the group areas numbering obtained.

4. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network as claimed in claim 1, is characterized in that: in step (4), calculates the averaged feature vector of each grouping according to formula (6),

${\mathrm{\&sigma;}}^g=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{x}_j^g---\left(6\right)$

${\mathrm{\&sigma;}}^g={[{\mathrm{\&sigma;}}_R^g,{\mathrm{\&sigma;}}_G^g,{\mathrm{\&sigma;}}_B^g]}^T$

Wherein σ ^gfor the mean eigenvalue of RGB color space, be respectively RGB component mean value, M is the pixel number of the grouping being numbered g, and g is the numbering of grouping, and j is neighbor subscript.

5. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network as claimed in claim 1, it is characterized in that: in step (6), for arbitrary neuron u do not connected, be connected in immediate neighboring region according to formula (7)

$j_u=\underset{j}{\arg }\left(\underset{j=1,...,8}{\min }\right||x^u-{\mathrm{\&sigma;}}^{g_j}|\left|\right)---\left(7\right)$

Wherein, u is for connecting pixel, x ^uthe proper vector of this pixel, that neighboring region is numbered g _javeraged feature vector; All neurons that do not connect are operated, and the difference not connecting neuron and neighboring region must be less than defined threshold simultaneously, and upgrade threshold value:

θ _i＝θ _i+Δθ _i(8)

Wherein, θ _ifor defined threshold, Δ θ _ifor threshold delta.

6. the Segmentation of Color Image of hyperchannel Pulse Coupled Neural Network as claimed in claim 1, is characterized in that: in step (7), in region merging technique process, using region distance and area size as two judgment condition:

(1) if the manhatton distance of two adjacent areas is less than predetermined threshold value, be a region by two region merging technique, and zoning average again;

(2) if area size is less than predetermined threshold value, this region is incorporated to the adjacent area minimum with its manhatton distance.