Pulse image segmentation method based on principal component analysis and cluster analysis

Abstract

The present application relates to the field of image processing of bionic image sensor, for realizing dimension reduction of visual pulse data and clustering analysis of similar trigger pulse rate pixel unit, thereby realizing image segmentation of scene. For this purpose, the technical scheme adopted by the present application is a pulse image segmentation method based on principal component analysis and clustering analysis, which independently converts optical signal into bit stream for each pixel in the image sensor, converts the bit stream IF model into pulse response, converts the data matrix to low-dimensional space by using the principal component analysis method PCA, realizes dimension reduction processing of pulse response data; according to the distance of different pulse responses to the clustering center, the clustering analysis is used as the criterion for distinguishing the category of the data, and the pulse sequences with similar trigger pulse rates are clustered, thereby realizing pulse image segmentation. The present application is mainly applied to the design and manufacturing application occasions of bionic image sensor.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing of bionic image sensors, in particular, to the field of dynamic perception and pulse data processing when image information in a scene changes rapidly, namely a pulse image segmentation algorithm based on principal component analysis and cluster analysis. BACKGROUND

[0002] How to make the rapidly fleeting object be able to be fixed and clearly displayed is the difficulty and pain point of the current high-end application of visual chip and system industry. The pulse visual model breaks the basic way of recording visual information by using images, changes a group of photons into a bit, and directly converts a light change process into a bit stream, realizing super high-speed photography.

[0003] In the pulse pixel, the light intensity information is converted into a voltage signal, and the process is completed by accumulating photo-generated charges. The accumulation rate of charges is different under different light intensity. When the photo-generated charges accumulate to a certain number, causing the voltage to reach a set threshold, the comparator completes the generation of the pulse signal. At this time, the photodiode is reset, and the accumulated photo-generated charges are emptied. The pixel enters the next integration stage. The analog visual neural photoreception mechanism uses an electric pulse pixel array to express external environmental light information, which fundamentally breaks through the bottleneck of the huge data volume, transmission speed and processing speed of existing image sensors, and fully surpasses the performance of human eye vision.

[0004] One way to reduce image information redundancy and break through the bottleneck of data transmission and processing is to simulate the intelligent sensing mechanism of biological vision, combine photoelectric sensing with neuromorphic computing, and perform pixel-level array parallel light intensity pulse signal from the perception end. However, traditional image processing algorithms are difficult to be directly applied to pulse signal processing. In addition, the high dimension of visual pulse data makes it difficult to directly perform image segmentation. Dimension reduction is a kind of pre-processing method for high-dimensional feature data, which retains the most important features of high-dimensional data, so as to realize the purpose of improving data processing speed. The present application reduces the dimension of visual pulse data based on principal component analysis and cluster analysis, and clusters the pixel units with similar trigger pulse rates, so as to realize image segmentation of the scene. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the present application aims to provide a pulse image segmentation algorithm based on principal component analysis and clustering analysis, which reduces the dimension of visual pulse data and analyzes the pixel units with similar trigger pulse rates, so as to realize image segmentation of a scene. To this end, the technical scheme adopted by the present application is a pulse image segmentation method based on principal component analysis and clustering analysis, which converts the optical signal into a bit stream independently by each pixel in the image sensor, converts the bit stream IF (Integrate and Fire) model into a pulse response, converts the data matrix into a low-dimensional space by using the principal component analysis method PCA (Principal Component Analysis), and realizes the dimension reduction processing of the pulse response data; the clustering analysis is used to take the distance from different pulse responses to the clustering center as the criterion for distinguishing the category to which the data belongs, and the pulse sequences with similar trigger pulse rates are clustered, so as to realize pulse image segmentation.

[0006] The specific steps of converting the data matrix into a low-dimensional space by using the principal component analysis method PCA are as follows:

[0007] The neuron model simulating the pulse coding process of biological neurons is called the IF model, the pulse response function is shown in formula (1), the pulse pixel converts the light intensity information into a time-domain pulse signal similar to a visual pulse, quantizes the scene information with low data volume and high transmission volume, and the output single-bit pulse signal is shown in formula (2):

[0008]

[0009]

[0010] Where τ is the decay time constant, t i k represents the kth visual pulse of the ith pulse sequence;

[0011] The pulse pixel quantizes the light intensity information in the form of a pulse, and converts the pulse sequence into a pulse response function as shown in formula (3):

[0012]

[0013] Input: pulse response function X0={X i}={x1,x2,x3,…,x n}, which is reduced to k dimensions:

[0014] (1) The visual pulse data set X0 is decentered to obtain X, that is, each pulse sequence is subtracted by the average value;

[0015] (2) The covariance matrix XX T, the eigenvalue ev and the eigenvector W are solved based on an eigenvalue decomposition method;

[0016] (3) The eigenvalues ev are sorted from large to small, and the cumulative contribution rate is calculated as shown in formula (4). The number of principal components r is determined according to the cumulative contribution rate reaching more than 80%.

[0017] (4) The pulse response sample principal component T is obtained based on the corresponding r eigenvectors i i T X, and the data is converted into a new space constructed by r eigenvectors:

[0018]

[0019] The PCA algorithm maps the p-dimensional features to r-dimensional features. The r-dimensional features are new orthogonal features also known as principal components, which are reconstructed r-dimensional features based on the original p-dimensional features.

[0020] The specific steps of clustering the pulse sequences with similar trigger pulse rates are as follows:

[0021] Input: pulse response sample principal component Y = {Y i} = {y1, y2, y3, …, y r}, which needs to be clustered into m classes, m < r;

[0022] (1) Randomly extract m objects from the pulse response sample principal component Y as initial cluster centers;

[0023] (2) Calculate the Euclidean distance of all objects Y i to the m cluster centers respectively, and compare them to assign the object Y i to the class with the smallest distance;

[0024] (3) According to the initial classification obtained in step (2), calculate the mean vector of each class as the new cluster center of each class;

[0025] (4) According to the new cluster center, repeat steps (2) and (3) until the iteration number is satisfied.

[0026] The characteristics and benefits of the present application are:

[0027] The present application proposes a pulse image segmentation algorithm based on principal component analysis and clustering analysis according to the data processing requirements in the high-speed imaging field. The dimension reduction of pulse data reduces the pulse data processing delay and realizes high-speed image processing of dynamic scene changes. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 Pulse image segmentation algorithm flow based on principal component analysis and clustering analysis.​

[0029] Figure 2 Impulse response function.

[0030] Figure 3 Cumulative contribution rate of principal component analysis and the number of principal components.

[0031] Figure 4 Different categories of visual impulse schematic diagram.

[0032] Figure 5 Dynamic scene visual impulse imaging and clustering analysis schematic diagram. In the figure:

[0033] a: visual impulse visualization in dynamic scene;

[0034] b: visual impulse image segmentation in dynamic scene. DETAILED DESCRIPTION

[0035] The impulse camera simulates the encoding principle of primate retina, each pixel independently converts the light signal into a bit stream, converts the bit stream IF model into an impulse response, and records the spatiotemporal variation process of light. The principal component analysis method PCA converts the data matrix to a low-dimensional space, realizing the dimensionality reduction processing of the impulse response data. The clustering analysis takes the distance of different impulse responses to the clustering center as the criterion for distinguishing the category to which the data belongs, and clusters the impulse sequences with similar trigger pulse rates, thereby realizing impulse image segmentation. The impulse image segmentation algorithm based on principal component analysis and clustering analysis is shown in Figure 1 .

[0036] In biological vision, light is incident on the retina through the pupil, and photoreceptor cells convert incident light into electrical signals. The membrane of neurons accumulates electric charge, and the voltage on the cell membrane fluctuates. Once the membrane potential exceeds a certain threshold, a visual impulse will be triggered. Neuron dynamics can be assumed to be a summation process combined with a mechanism that triggers an action potential above a critical voltage. The neuron model that simulates the biological neuron pulse encoding process is called IF model, and the impulse response function is shown in formula (1). Unlike traditional active pixel structures, the impulse pixel converts light intensity information into time-domain pulse signals similar to visual impulses, quantizes scene information with low data volume and high transmission rate, and the output single-bit pulse signal is shown in formula (2).

[0037]

[0038]

[0039] where τ is the decay time constant, t i k represents the kth visual impulse of the ith impulse sequence.

[0040] The pulse pixel quantifies the light intensity information in the form of pulse electrical signal. In order to facilitate the multivariate statistical analysis of the pulse information, the pulse sequence is converted into a pulse response function as shown in equation (3).

[0041]

[0042] Based on the pulse response function converted from the pulse sequence, the PCA algorithm of eigenvalue decomposition of the covariance matrix is performed, and the specific steps are as follows.

[0043] Input: pulse response function X0={X i}={x1,x2,x3,…,x n} needs to be reduced to k dimensions.

[0044] (1) The visual pulse data set X0is decentralized to obtain X, that is, each pulse sequence is subtracted by the average value.

[0045] (2) Calculate the covariance matrix XX T of the visual pulse data set X, and solve the eigenvalue evand the eigenvector Wbased on the eigenvalue decomposition method.

[0046] (3) Sort the eigenvalues evfrom large to small, calculate the cumulative contribution rate as shown in equation (4), and determine the number of principal components raccording to the cumulative contribution rate reaching more than 80%.

[0047] (4) Based on the corresponding r eigenvectors, the pulse response sample principal component Y i =W i T X, the data is converted into a new space constructed by r eigenvectors.

[0048]

[0049] The PCA algorithm maps the p-dimensional features to r-dimensional features. The r-dimensional features are new orthogonal features also known as principal components, which are r-dimensional features reconstructed on the basis of the original p-dimensional features. The specific steps of clustering analysis based on the reduced dimension data obtained by the PCA algorithm are as follows.

[0050] Input: pulse response sample principal component Y={Y i}={y1,y2,y3,…,y r}, which needs to be clustered into m(m

[0051] (1) Randomly extract m objects from the pulse response sample principal component Y as initial cluster centers.

[0052] (2) Calculate the Euclidean distance of all objects Y i to the m cluster centers respectively, and compare them to obtain the object Y iThe class with the smallest distance is selected.

[0053] (3) According to the initial classification obtained in step (2), the mean vector of each class is calculated as the new cluster center of each class.

[0054] (4) According to the new cluster center, steps (2) and (3) are repeated until the number of iterations is satisfied.

[0055] The application is based on principal component analysis and cluster analysis, and realizes data dimension reduction and clustering of similar trigger pulse rate pixel units of visual pulses output by a super high-speed pulse camera, so as to realize simple image segmentation of a scene. When the decay time constant τ is 0.0004, the pulse response function schematic diagram of three pulse sequences is as shown in Figure 2 The relationship between the cumulative contribution rate of principal component analysis and the number of principal components is as shown in Figure 3 The number of principal components r is determined to be 2 according to the cumulative contribution rate of more than 80%. Based on cluster analysis, different objects in a dynamic scene are clustered and classified into three classes. The visual pulse schematic diagram of different classes is as shown in Figure 4 The dynamic scene visual pulse imaging and cluster analysis schematic diagram is as shown in Figure 5

[0056] The above is only a specific embodiment of the application, but the protection scope of the application is not limited to this. Any changes or replacements within the technical range disclosed by the application can be easily thought by those skilled in the art, and should be covered within the protection scope of the application.​

Claims

1. A pulse image segmentation method based on principal component analysis and cluster analysis, characterized by, The bit stream is converted into a pulse response by converting the light signal into the bit stream independently by each pixel in the image sensor, the data matrix is converted into a low-dimensional space by a principal component analysis method PCA, and dimensionality reduction processing of the pulse response data is realized; The pulse image is segmented by clustering similar trigger pulse rates of the pulse sequence according to the distance of different pulse responses to the cluster center as a criterion for distinguishing the category to which the data belongs; The specific steps of converting the data matrix into a low-dimensional space by the principal component analysis method PCA are as follows: The neuron model simulating the biological neuron pulse coding process is called an IF model, the pulse response function is shown in formula (1), the pulse pixel converts the light intensity information into a time-domain pulse signal similar to a visual pulse, and the scene information is quantized at a low data rate and high transmission rate, and the output single-bit pulse signal is shown in formula (2): where τ is the decay time constant, t i k represents the kth visual pulse of the ith pulse train; The pulse pixel quantizes the light intensity information in the form of a pulse, and the pulse sequence is converted into a pulse response function as shown in formula (3): Input: Impulse response function X0 = {X i} = {x1, x2, x3,..., x n} is reduced to k dimensions: (1) The visual pulse data set X0 is decentralized to obtain X, that is, each pulse sequence is subtracted from the average value; (2) computing a covariance matrix XX of the visual pulse data set X T solving the eigenvalue ev and the eigenvector W based on an eigenvalue decomposition method; (3) The characteristic values ev are sorted from large to small, the cumulative contribution rate is calculated as shown in formula (4), and the number of principal components r is determined according to the cumulative contribution rate reaching more than 80%; (4) Get the principal component T of the impulse response sample based on the corresponding r eigenvectors i = W i T X, convert the data to the new space constructed by the r eigenvectors The PCA algorithm maps the p-dimensional feature to r-dimensional, and the r-dimensional is a new orthogonal feature also called principal component, which is a r-dimensional feature reconstructed on the basis of the original p-dimensional feature.

2. The pulse image segmentation method based on principal component analysis and cluster analysis according to claim 1, wherein, The specific steps of clustering the pulse sequence with similar trigger pulse rates are as follows: Input: Impulse response samples principal components Y = {Y i} = {y1, y2, y3,..., y r} that need to be clustered into m classes, m < r. (1) Randomly extract m objects from the principal components Y of the pulse response sample as initial cluster centers; (2) Calculate the Euclidean distance of all objects Y i to m cluster centers, compare with each other and attribute object Y i to the class with the smallest distance. (3) According to the initial classification obtained in step (2), the mean vector of each class is calculated as the new cluster center of each class; (4) According to the new cluster center, steps (2) and (3) are repeated until the number of iterations is satisfied.

Images