CN110276776B - Adaptive target detection method based on SPCNN - Google Patents

Adaptive target detection method based on SPCNN Download PDF

Info

Publication number
CN110276776B
CN110276776B CN201910521356.2A CN201910521356A CN110276776B CN 110276776 B CN110276776 B CN 110276776B CN 201910521356 A CN201910521356 A CN 201910521356A CN 110276776 B CN110276776 B CN 110276776B
Authority
CN
China
Prior art keywords
image
spcnn
adaptive
iteration
target
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910521356.2A
Other languages
Chinese (zh)
Other versions
CN110276776A (en
Inventor
周肃
宋勇
郭拯坤
张大勇
赵宇飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Borui Weixin Technology Co ltd
Beijing Institute of Technology BIT
Original Assignee
Beijing Borui Weixin Technology Co ltd
Beijing Institute of Technology BIT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Borui Weixin Technology Co ltd, Beijing Institute of Technology BIT filed Critical Beijing Borui Weixin Technology Co ltd
Priority to CN201910521356.2A priority Critical patent/CN110276776B/en
Publication of CN110276776A publication Critical patent/CN110276776A/en
Application granted granted Critical
Publication of CN110276776B publication Critical patent/CN110276776B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Abstract

The invention discloses a self-adaptive target detection method based on SPCNN, belonging to the technical field of computer vision. The implementation method of the invention comprises the following steps: calculating static attribute parameters of the image; deducing a theoretical formula according to Stevens's law, and calculating a threshold decay time constant alphaeMake the threshold decay time constant alphaeThe self-adaptive setting can be realized according to the overall gray characteristic of the target image; based on a self-adaptive side inhibition mechanism, improving an inhibition coefficient calculation model by using a hyperbolic tangent function, and calculating a link weight matrix of each pixel point by using the inhibition coefficient calculation model; adaptively setting image input parameters in an intact SPCNN, continuously iterating, generating a binarization segmentation result, and extracting a candidate target; based on a quick connection mechanism in neuron synchronization, the automatic output of the optimal segmentation result is realized by calculating the similarity of adjacent iteration segmentation results and searching the maximum value of the similarity by combining a gray image criterion, and meanwhile, the iteration is automatically controlled, so that the efficiency and the intelligence of the target detection method are improved.

Description

Adaptive target detection method based on SPCNN
Technical Field
The invention relates to a self-adaptive target detection method, in particular to a self-adaptive target detection method based on a Simplified Pulse Coupled Neural Network (SPCNN), and belongs to the technical field of computer vision.
Background
Pulse Coupled Neural Network (PCNN) belongs to a novel Neural Network model of the third generation artificial Neural Network. The Pulse Coupled Neural Network (PCNN) is established by inspiring the activity of neurons in the main visual region V1 of the neocortex of the human brain, and has the advantages of translation, rotation, scale invariance, good noise resistance, no need of training and the like in the aspect of processing digital images. The attributes of the PCNN for the image processing mechanism are divided into completely new three dimensions, the first dimension specifies the time matrix of the PCNN, the second dimension captures the emission rate of the PCNN, and the third dimension is the synchronization of the PCNN.
At present, the pulse coupling neural network applied to the fields of image segmentation, target detection and the like mainly has two problems: (1) the segmentation effect of the conventional PCNN model depends largely on the parameter settings of the PCNN. The PCNN model has more parameters and can be estimated only through manual adjustment or a large number of experiments, so that the applicability and the mobility of the algorithm are poor; (2) each iteration of the PCNN network generates a binary output, which usually requires manual selection of an optimal result, which is not conducive to its engineering application.
In view of the above problems, methods of image segmentation, object detection, and the like based on a Simplified Pulse Coupled Neural Network (SPCNN) have been proposed in succession. The SPCNN simplifies the conventional PCNN network, and can obtain better results while reducing the complexity of the network and reducing parameters. The method mainly comprises the following algorithms: (1) MA et al propose a new image segmentation algorithm for automatic control iteration by combining the minimum cross entropy criterion with respect to the conventional PCNN threshold segmentation mechanism. The method can realize automatic judgment of the PCNN iteration times and automatic selection of optimal output. However, the cross entropy calculation in this method is complex and time-consuming. Meanwhile, for a target under a complex background, certain errors exist in information entropy measurement before and after image segmentation, so that the accuracy of a detection result is difficult to guarantee. (2) Zhan et al put forward a new spike cortical neural network model based on the conventional PCNN model, and put forward that the time matrix of the model can be regarded as the subjective stimulation intensity sense of human beings; chen et al propose a new automatic parameter setting method based on the spiking cortical neural network model. The method can realize the self-setting of most parameters according to the static attributes of the image. However, the individual parameters in this method still need to be manually set, and meanwhile, the calculation method of some parameters has certain defects. (3) MA et al propose a method for establishing a link weight matrix according to local gray level correlation and Euclidean distance of an image for the link weight matrix of the SPCNN model, and simultaneously propose a new method for automatically controlling iteration by combining a minimum variance ratio criterion. The method can achieve a better segmentation effect under most conditions. However, the minimum variance ratio in the method is more complicated to calculate and takes longer, so that the method has the defects of difficult application and the like.
Aiming at the problems, on the basis of the research of a human brain visual mechanism, a new target detection method based on SPCNN parameter self-adaptation and iterative self-adaptation is researched, and the method has important significance and wide application prospect for solving the target detection problem under the complex background.
Disclosure of Invention
Aiming at the problems that the SPCNN needs to manually set parameters and manually select optimal output, and the existing method is complex in calculation, long in time consumption, poor in robustness and the like. The invention discloses a self-adaptive target detection method based on SPCNN, which aims to solve the technical problems that: the method can adaptively set the parameters of the SPCNN and automatically control iteration, thereby improving the robustness, the anti-interference capability and the intelligence of the target detection method.
The purpose of the invention is realized by the following technical scheme.
The invention discloses a self-adaptive target detection method based on SPCNN. Firstly, calculating static attribute parameters of the image, including image gray level mean value mu and optimal histogram threshold value TotsuAnd the image standard deviation σ; secondly, a theoretical formula is deduced according to Stevens's law, and a threshold decay time constant alpha is calculatedeMake the threshold decay time constant alphaeCan be based on the target imageThe overall gray scale characteristics are set in a self-adaptive manner, so that the reliability and robustness of the target detection method are improved; thirdly, based on a self-adaptive side suppression mechanism, improving a suppression coefficient calculation model by using a hyperbolic tangent function, and calculating a link weight matrix of each pixel point by using the suppression coefficient calculation model, so that the anti-interference capability of the target detection method is improved; then, adaptively setting image input parameters in an intact SPCNN, continuously iterating, generating a binarization segmentation result, and extracting a candidate target; and finally, based on a quick connection mechanism in neuron synchronization, combining a gray level image criterion, automatically outputting the optimal segmentation result by calculating the similarity of adjacent iteration segmentation results and searching a maximum value of the similarity, and automatically controlling iteration at the same time, thereby improving the efficiency and the intelligence of the target detection method.
The invention discloses a self-adaptive target detection method based on SPCNN, which comprises the following steps:
step 1: and calculating the static attribute parameters of the image.
After the image is subjected to graying preprocessing, the integral gray level mean value of the image is calculated, and the mean value is used as a calculation factor of a threshold value attenuation time constant in the SPCNN.
And normalizing the image gray matrix, and searching the highest gray value in the background, wherein the highest gray value is used as a calculation factor of a link factor and a threshold amplification constant in the SPCNN. The highest gray value in the background is the best histogram threshold.
And meanwhile, calculating the standard deviation of the normalized gray level image, wherein the standard deviation is used as an exponential decay coefficient of the SPCNN internal activity item.
The image static attribute parameters comprise an image gray level mean value, an optimal histogram threshold value and a gray level image standard deviation.
Preferably, Otsu's method is used to find the highest gray value in the background in step 1.
Step 2: deducing a theoretical formula according to Stevens's law, and calculating a threshold decay time constant alphaeMake the threshold decay time constant alphaeCan be self-adaptively set according to the overall gray-scale characteristics of the target image,and further, the reliability and robustness of the target detection method are improved.
In the prior art, the manually set threshold decay time constant alpha is based on Stevens's laweThe error is large, and the non-target pixel point is possibly misfired. From the subjective angular analysis of human vision, the threshold decay time constant αeAnd an inverse proportional relation exists between the image gray level mean value mu and the inverse proportional relation is as follows: when the image gray level mean value mu is smaller, the applicable threshold decay time constant alphaeShould be greater than the empirical value; when the image gray level mean value mu is larger, the applicable threshold attenuation time constant alphaeShould be much smaller than the empirical value. From an objective analysis of image gray scale, the threshold decay time constant αeAnd an inverse proportional relation also exists with the image gray level mean value mu, and the inverse proportional relation is as follows: in most infrared images the target is of a higher gray level, so that a larger image mean μ indicates that the target and background gray levels are closer, requiring less threshold attenuation for finer segmentation, and hence a required threshold attenuation time constant αeSmaller; conversely, when the image gray mean value mu is smaller, the background gray is lower and the target is more prominent, so that the accurate segmentation can be completed by larger threshold attenuation, and the required threshold attenuation time constant alpha iseAnd is larger.
According to Stevens's law, the subjective feeling quantity and the objective stimulation quantity of the light intensity form a power function relationship with the exponent of 0.5, and the deduced threshold decay time constant alphaeThe mathematical description of the image gray level mean μ is shown in equation (1):
Figure BDA0002096783230000041
in the formula, μ represents a grayscale mean value of an image.
Adaptively calculating a threshold decay time constant alpha according to equation (1)eNamely, the threshold decay time constant alpha is set adaptively according to the overall gray level characteristics of the target imagee。
And step 3: based on a self-adaptive side inhibition mechanism, an inhibition coefficient calculation model is improved by utilizing a hyperbolic tangent function, and a link weight matrix of each pixel point is calculated by using the inhibition coefficient calculation model, so that the anti-interference capability of the target detection method is improved.
The advantages of the adaptive side suppression mechanism include strong background suppression, and enhanced contrast between the target and the background, so that the target can be better highlighted and the texture details of the target are retained.
Based on an Adaptive Lateral Inhibition mechanism (ALI), the Inhibition coefficient calculation model is improved by utilizing a hyperbolic tangent function to obtain an Adaptive Inhibition coefficient calculation model, the Adaptive Inhibition coefficient calculation model is characterized by the hyperbolic tangent function with a definition domain of [0,3], and the corresponding curve form and value domain can simulate the Lateral Inhibition mechanism more appropriately. The self-adaptive suppression coefficient calculation model is used for replacing a conventional link weight matrix, distance factors and gray difference factors among pixels are considered in calculation of elements in the matrix, and therefore the link weight matrix corresponding to each pixel point can be calculated in a self-adaptive mode according to information contained in an image, and the link weight matrix is as shown in a formula (2):
Wijkl＝1-tanh[I(i,j)×dij,kl×10-2] (2)
in the formula, WijklRepresenting the link weight matrix, I (I, j) representing the gray level of the pixel points of the input image, dij,klRepresenting the euclidean distance between two pixels.
And 4, step 4: and (4) adaptively setting image input parameters in the intact SPCNN, continuously iterating, generating a binarization segmentation result, and extracting a candidate target.
In the simplified pulse coupling neural network SPCNN model, discrete mathematical description of single neuron function realization is shown in formulas (3) to (5):
Figure BDA0002096783230000042
Figure BDA0002096783230000043
Figure BDA0002096783230000044
in the formula of Uij[n]Is an internal activity term of a neuron, alphafDecay time constant for internal activity, SijFor an externally input stimulus signal, i.e., a gray distribution of an input image, β is a connection strength coefficient between synapses, VLIs set as a connection amplitude constant, WijklTo link the weight matrix, Yij[n-1]Is the output of the neuron at the (n-1) th iteration; y isij[n]Output of neurons at the nth iteration, Eij[n-1]The dynamic threshold of the neuron at the (n-1) th iteration; alpha is alphaeIs a threshold decay time constant, VEIs the dynamic threshold amplitude. When the neuron internal activity item Uij[n]Greater than dynamic threshold Eij[n-1]And if not, the neuron is not excited and does not generate pulse output.
Input image SijAnd as external input stimulation, inputting the image into the SPCNN for operation, and determining whether to ignite each pixel according to the gray distribution of each pixel and the neighborhood of the pixel by the network so as to obtain a segmented binary image and extract a candidate target.
And 5: based on a quick connection mechanism in neuron synchronization, the automatic output of the optimal segmentation result is realized by calculating the similarity of adjacent iteration segmentation results and searching the maximum value of the similarity by combining a gray image criterion, and meanwhile, the iteration is automatically controlled, so that the efficiency and the intelligence of the target detection method are improved.
The mechanism of fast connection in neuron synchronization, referred to as the mechanism of fast connection for short, is faster than other normal neurons synchronizing with the same stimulus. Grayscale images typically meet the following criteria: (1) the gray level similarity between the target point and the target point is high; (2) the gray level similarity between the target point and the background point is low; (3) the gray level similarity between the background point and the background point is low.
According to a quick connection mechanism and in combination with a criterion (1), starting from the ignition of a target pixel, the target pixel is continuously ignited, the detail information of the target is continuously perfected, the background is always in a suppressed state, and the similarity of the segmentation results of two adjacent iterations is gradually increased in the stage; and when the background pixel points are continuously ignited, the similarity of the segmentation results of two adjacent iterations is suddenly reduced according to the criteria (2) and (3). Therefore, in the whole iteration process, the image similarity of the adjacent segmentation results is in the relationship of increasing and then decreasing. And representing the segmentation images in each iteration by adopting a Hash value, and further representing the similarity by utilizing the Hamming distance between the segmentation images, wherein the smaller the Hamming distance is, the higher the similarity is. Then, in the iteration process, the similarity peak value of adjacent segmentation results, namely the minimum value of the Hamming distance, is searched, the segmentation image corresponding to the iteration is the optimal output result, and the iteration is judged to be ended.
The step 5 is realized as follows:
step 5.1: and calculating Hash values of output results of two adjacent iterations, namely H (i) and H (j), (i-1, 2, …, N, j-i + 1). Using h (i) and h (j) to calculate the hamming distance between them, which is denoted as hds (N), (N ═ 2, …, N).
Step 5.2: the smaller the Hds (n), the higher the similarity between the two adjacent iteration segmentation results, and then in the previous n iterations, the minimum value of Hds (n) is found.
Step 5.3: calculating the gray level average value of the segmentation result in each iteration, and recording the gray level average value as mean (N), (N is 1,2, …, N), wherein in order to avoid finding that hds (N) is a local minimum value, an additional condition mean (N) > preset threshold value needs to be added. When the two conditions are met, the optimal output result is obtained, and the judgment iteration is finished at the same time, so that the automatic selection of the optimal output result is completed, and the efficiency and the intellectualization of the target detection method are improved.
Preferably, the preset threshold value in step 5.3 is set to ensure that the detected target is meaningful, preferably 0.0095-0.025, and more preferably 0.01.
Has the advantages that:
1. and the robustness is strong. The invention discloses a self-adaptive target detection method based on SPCNN, and provides a self-adaptive calculation method of a threshold attenuation constant and a link weight matrix. The two methods do not need to adjust parameters according to empirical values, but set each parameter in a self-adaptive manner according to the static attribute of the image, so that the robustness of the method is effectively improved.
2. The anti-interference capability is strong. The invention discloses an adaptive target detection method based on SPCNN, which improves an SPCNN model based on Stevens law and a side suppression mechanism, can enhance the contrast ratio of a target and a background and completely reserve the texture details of the target by improving the SPCNN model, and can effectively suppress clutter interference in a complex background, thereby improving the target detection precision.
3. Intelligence is high-efficient. The invention discloses an adaptive target detection method based on SPCNN, which is based on a quick connection mechanism in neuron synchronization, combines a gray image criterion, realizes automatic output of an optimal segmentation result by calculating the similarity of adjacent iterative segmentation results and searching a maximum value of the similarity, and simultaneously automatically controls iteration. The method is used as a selection criterion of the optimal output result, has simple calculation and high efficiency, and effectively improves the intelligent degree of the method.
Drawings
FIG. 1 is a flow chart of an adaptive target detection method based on SPCNN disclosed in the present invention;
FIG. 2 is a comparison of the quality of the detection results of the target detection method of the present invention and the conventional target detection and target segmentation methods. FIGS. 2(a) and (b) are comparison of target detection results in a low contrast environment; FIGS. 2(c) and (d) are comparison of target detection results in complex background; FIG. 2(e) is a comparison of the results of target detection with noise interference; fig. 2(f) is a comparison of the detection results of infrared small and weak targets.
Detailed Description
For a better understanding of the objects and advantages of the present invention, reference should be made to the following detailed description taken in conjunction with the accompanying drawings and examples.
Example 1:
the overall process of the adaptive target detection method based on SPCNN disclosed in this embodiment is shown in fig. 1, and the specific implementation steps are as follows:
step 1: and calculating static attribute parameters of the image and partial parameters in the SPCNN model.
After the image is subjected to graying preprocessing, the integral gray level mean value of the image is calculated, and the mean value is used as a calculation factor of a threshold value attenuation time constant in the SPCNN model.
Then, the image gray matrix is normalized, the maximum gray value of the target pixel area is calculated, the best histogram threshold is obtained by using the Otsu method, and the standard deviation of the normalized gray image is calculated.
The above three parameter values will be used as calculation factors of the parameters in the SPCNN model.
According to the existing method, the parameter values obtained by preprocessing are used for carrying out self-adaptive setting on the internal activity exponential decay constant, the connection coefficient and the dynamic threshold amplitude.
Step 2: deducing a theoretical formula according to Stevens's law, and calculating a threshold decay time constant alphaeMake the threshold decay time constant alphaeThe self-adaptive setting can be realized according to the overall gray scale characteristics of the target image.
Stevens's law states that the magnitude of the subjective sensory quantity is proportional to the power of the objective stimulus quantity, i.e., the psychological quantity is a power function of the physical quantity. The expression is as follows:
S＝K×In (6)
wherein S is a psychological quantity, K is a constant, I is a physical quantity, and n is different due to different senses. When used to express the property of light intensity, n is 0.5.
The human eye vision mechanism is introduced into the threshold decay time constant alphaeIn the calculation model of (2), alpha is obtained by analysiseAnd is inversely proportional to the half power of the image gray level mean value mu. The expression is as follows:
Figure BDA0002096783230000071
in the formula, μ represents a grayscale mean value of an image.
And (3) substituting the image integral gray average value calculated in the step (1) into a formula (7) to obtain a corresponding threshold attenuation time constant.
And step 3: based on a self-adaptive side inhibition mechanism, an inhibition coefficient calculation model is improved by utilizing a hyperbolic tangent function, and a link weight matrix of each pixel point is calculated by using the inhibition coefficient calculation model.
A side inhibition mechanism is simulated by using a hyperbolic tangent function with a definition domain of [0,3], and a conventional link weight matrix is replaced by a self-adaptive inhibition coefficient calculation model. The distance factor and the gray difference factor between pixels are simultaneously considered in the calculation of each element in the matrix, so that the link weight matrix corresponding to each pixel point can be calculated in a self-adaptive manner according to the information contained in the image. The mathematical description is shown in equation (8):
Wijkl＝1-tanh[I(i,j)×dij,kl×10-2] (8)
in the formula, WijklRepresenting the link weight matrix, I (I, j) representing the gray level of the pixel points of the input image, dij,klRepresenting the euclidean distance between two pixels.
After the balance between the information quantity and the calculation quantity, a link weight matrix with the neighborhood size of 3 multiplied by 3 is selected. Meanwhile, the gray value of the image with the gray level range of 0-255 needs to be mapped to [0,3]]Within the domain of (3 x 3), the pixel pitch is 1 or
Figure BDA0002096783230000081
Therefore, the value range of the elements in the link weight matrix is [0,1 ]]。
And 4, step 4: and (4) adaptively setting image input parameters in the intact SPCNN, continuously iterating, generating a binarization segmentation result, and extracting a candidate target.
And separating the possible target from the background clutter and the noise by using an SPCNN method, and extracting a candidate target.
In the SPCNN model, discrete mathematical descriptions of single neuron function implementation are shown in equations (9) to (11):
Figure BDA0002096783230000082
Figure BDA0002096783230000083
Figure BDA0002096783230000084
in the formula of Uij[n]Is an internal activity term of a neuron, alphafDecay time constant for internal activity, SijFor an externally input stimulus signal, i.e., a gray distribution of an input image, β is a connection strength coefficient between synapses, VLIs set to constant 1, WijklTo link the weight matrix, Yij[n-1]Is the output of the neuron at the (n-1) th iteration; y isij[n]Output of neurons at the nth iteration, Eij[n-1]The dynamic threshold of the neuron at the (n-1) th iteration; alpha is alphaeIs a threshold decay time constant, VEIs the dynamic threshold amplitude. When the neuron internal activity item Uij[n]Greater than dynamic threshold Eij[n-1]And if not, the neuron is not excited and does not generate pulse output.
Will input the image SijAnd as external input stimulation, inputting the image into an SPCNN network for operation, and determining whether to ignite the pixel by the SPCNN model according to the gray distribution of each pixel and the surrounding area thereof, so as to obtain a segmented binary image and extract a candidate target.
The specific implementation of step 4 is performed according to the following steps, and the specific process is described by taking the 1 st iteration as an example:
step 4.1 link input fields.
In the first iteration of the SPCNN network, the neuron internal activity expressions of the linked domain are as follows:
Figure BDA0002096783230000091
in the formula, SijRepresents an input image, [1 ]]Denotes iteration 1, Uij[1]Is the main input to the (i, j) th neuron.
Before the SPCNN network is iterated for the first time, because no external stimulus exists, all the neurons are in an unexcited state, and then the initial internal activity item U of the (i, j) th neuronij[0]And pulse output Yij[0]Are all zero. The above equation can therefore be simplified to:
Uij[1]＝Sij (13)
step 4.2 pulse generation domain.
In the pulse generation domain, an output signal is determined through comparison of the internal activity item and a dynamic threshold, when the internal activity item of the neuron is larger than the dynamic threshold, the neuron is excited to generate pulse output, otherwise, the neuron is not excited to generate the pulse output. In the first iteration, the dynamic threshold value Eij[0]Are all 0, as can be seen from formula (13), Uij[1]≥Eij[0]Therefore, the following are:
Yij[1]＝1 (14)
in the formula, Yij[1]Representing the pulse output of the first iteration. In the first iteration, all neurons are excited because the activity item inside the neuron is larger than the dynamic threshold, i.e. the pixel value is 1 in the binary output.
Step 4.3 threshold decay domain.
The dynamic threshold expression for the first iteration is as follows:
Figure BDA0002096783230000092
in the formula, dynamic threshold value Eij[0]Are all 0, neuron pulse output Yij[1]Are all 1, threshold decay time constant αeAnd dynamic threshold amplitude VEAre all set adaptively by the static attribute of the image. The above equation is therefore simplified to:
Eij[1]＝VE (16)
equation (16) shows that after the first iteration, the neuron dynamics thresholdAll values become VE. In the next iteration, the pixel is fired only if the neuron internal activity term is greater than the dynamic threshold, the neuron fires to produce a pulsed output. The SPCNN network determines whether to ignite the pixel through continuous iterative computation, so that a segmented binary image is obtained, and a candidate target is extracted.
And 5: based on a quick connection mechanism in neuron synchronization, the automatic output of the optimal segmentation result is realized by calculating the similarity of adjacent iteration segmentation results and searching the maximum value of the similarity by combining a gray image criterion, and the iteration is automatically controlled.
According to the fast connection mechanism, the similarity change of the segmentation results in adjacent iterations is basically found in the following three stages:
a. when the neuron corresponding to the target pixel point starts to ignite, the detail information of the target is gradually improved along with the iteration, and the neuron corresponding to the background pixel point is in a suppressed state because the gray scale similarity between the target and the background is low. The similarity of the segmentation results of two adjacent iterations assumes an increasing situation in this phase.
b. At this stage, the target is basically complete, the neurons corresponding to the background pixel points are not ignited, and the similarity of the segmentation results of two adjacent iterations reaches a maximum value.
c. When a background pixel point corresponds to a neuron, the neuron starts to ignite, because the gray scale similarity between the background pixel points is low, the neuron ignition is relatively random, and the similarity of the segmentation results of two adjacent iterations is suddenly reduced along with the iteration.
The algorithm flow is as follows:
(1) scaling the segmented image in the current iteration to 32 x 32;
(2) converting 32 × 32 to 1024-bit, each 8 bits being a hexadecimal value into a character string, and generating a Hash value;
(3) after the Hash value is generated, calculating the Hamming distance between the Hash value and the last Hash value, and judging the similarity of the two images;
(4) and when the Hamming distance has a global minimum value and the gray average value of the current segmentation result is greater than zero, the segmentation result corresponding to the iteration round is the optimal output, and the iteration is immediately judged to be finished.
The adaptive parameter values of the embodiments are shown in table 1, and the bold part in the table is the adaptive parameters and the number of iterations for obtaining the best output result according to the present invention.
Table 1 example adaptive parameter values and number of iterations
Figure BDA0002096783230000111
The above detailed description is intended to illustrate the objects, aspects and advantages of the present invention, and it should be understood that the above detailed description is only exemplary of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements and the like within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (9)

1. An adaptive target detection method based on SPCNN is characterized in that: comprises the following steps of (a) carrying out,
step 1: calculating static attribute parameters of the image;
step 2: deducing a theoretical formula according to Stevens's law, and calculating a threshold decay time constant alphaeMake the threshold decay time constant alphaeThe method can be set in a self-adaptive manner according to the overall gray level characteristics of the target image, so that the reliability and robustness of the target detection method are improved;
and step 3: based on a self-adaptive side inhibition mechanism, an inhibition coefficient calculation model is improved by utilizing a hyperbolic tangent function, and a link weight matrix of each pixel point is calculated by using the inhibition coefficient calculation model, so that the anti-interference capability of the target detection method is improved;
and 4, step 4: adaptively setting image input parameters in an intact SPCNN, continuously iterating, generating a binarization segmentation result, and extracting a candidate target;
and 5: based on a quick connection mechanism in neuron synchronization, combining a gray image criterion, automatically outputting an optimal segmentation result by calculating the similarity of adjacent iteration segmentation results and searching a maximum value of the similarity, and automatically controlling iteration at the same time, thereby improving the efficiency and the intellectualization of the target detection method;
wherein, the step 5 is realized by the following steps,
the fast connection mechanism in the neuron synchronization, which is referred to as the fast connection mechanism for short, is faster than other normal neurons in synchronization of neurons with the same stimulation; the grayscale image satisfies the following criteria: (1) the gray level similarity between the target point and the target point is high; (2) the gray level similarity between the target point and the background point is low; (3) the gray level similarity between the background point and the background point is low;
according to a quick connection mechanism and in combination with a criterion (1), starting from the ignition of a target pixel, the target pixel is continuously ignited, the detail information of the target is continuously perfected, the background is always in a suppressed state, and the similarity of the segmentation results of two adjacent iterations is gradually increased; when background pixel points are continuously ignited, the similarity of the segmentation results of two adjacent iterations is suddenly reduced according to the criteria (2) and (3); therefore, in the whole iteration process, the image similarity of adjacent segmentation results is in a relationship of increasing gradually and then decreasing gradually; representing the segmentation images in each iteration by adopting a Hash value, and further representing the similarity by utilizing the Hamming distance between the segmentation images, wherein the smaller the Hamming distance is, the higher the similarity is; then, in the iteration process, the similarity peak value of adjacent segmentation results, namely the minimum value of the Hamming distance, is searched, the segmentation image corresponding to the iteration is the optimal output result, and the iteration is judged to be ended.
2. The adaptive target detection method based on SPCNN of claim 1, wherein: the step 1 is realized by the method that,
after carrying out graying pretreatment on the image, calculating the integral gray average value of the image, wherein the average value is used as a calculation factor of a threshold value attenuation time constant in the SPCNN;
normalizing the image gray matrix, and searching the highest gray value in the background, wherein the highest gray value is used as a calculation factor of a link factor and a threshold amplification constant in the SPCNN; the highest gray value in the background is the optimal histogram threshold;
meanwhile, calculating the standard deviation of the normalized gray level image, wherein the standard deviation is used as an exponential decay coefficient of an SPCNN internal activity item;
the image static attribute parameters comprise an image gray level mean value, an optimal histogram threshold value and a gray level image standard deviation.
3. The adaptive target detection method based on SPCNN as claimed in claim 2, wherein: the step 2 is realized by the method that,
according to Stevens's law, the subjective feeling quantity and the objective stimulation quantity of the light intensity form a power function relationship with the exponent of 0.5, and the deduced threshold decay time constant alphaeThe mathematical description of the image gray level mean μ is shown in equation (1):
Figure FDA0002999580610000021
in the formula, μ represents a gray level mean value of an image;
adaptively calculating a threshold decay time constant alpha according to equation (1)eNamely, the threshold decay time constant alpha is set adaptively according to the overall gray level characteristics of the target imagee。
4. The adaptive target detection method based on SPCNN of claim 3, wherein: the step 3 is realized by the method that,
based on an adaptive side inhibition mechanism, improving an inhibition coefficient calculation model by utilizing a hyperbolic tangent function to obtain an adaptive inhibition coefficient calculation model, wherein the adaptive inhibition coefficient calculation model is characterized by the hyperbolic tangent function with a definition domain of [0,3], and the corresponding curve form and value domain can simulate the side inhibition mechanism more appropriately; the self-adaptive suppression coefficient calculation model is used for replacing a conventional link weight matrix, distance factors and gray difference factors among pixels are considered in calculation of elements in the matrix, and therefore the link weight matrix corresponding to each pixel point can be calculated in a self-adaptive mode according to information contained in an image, and the link weight matrix is as shown in a formula (2):
Wijkl＝1-tanh[I(i,j)×dij,kl×10-2] (2)
in the formula, WijklRepresenting the link weight matrix, I (I, j) representing the gray level of the pixel points of the input image, dij,klRepresenting the euclidean distance between two pixels.
5. The adaptive target detection method based on SPCNN of claim 4, wherein: step 4, the method is realized by the following steps,
in the simplified pulse coupling neural network SPCNN model, discrete mathematical description of single neuron function realization is shown in formulas (3) to (5):
Figure FDA0002999580610000031
Figure FDA0002999580610000032
Figure FDA0002999580610000033
in the formula of Uij[n]Is an internal activity term of a neuron, alphafDecay time constant for internal activity, SijFor an externally input stimulus signal, i.e., a gray distribution of an input image, β is a connection strength coefficient between synapses, VLIs set as a connection amplitude constant, WijklTo link the weight matrix, Yij[n-1]Is the output of the neuron at the (n-1) th iteration; y isij[n]Output of neurons at the nth iteration, Eij[n-1]The dynamic threshold of the neuron at the (n-1) th iteration; alpha is alphaeIs a threshold decay time constant, VEIs the dynamic threshold amplitude; spirit of the inventionChannel element internal activity item Uij[n]Greater than dynamic threshold Eij[n-1]When the neuron is excited to generate pulse output, otherwise, the neuron is not excited to generate pulse output;
input image SijAnd as external input stimulation, inputting the image into the SPCNN for operation, and determining whether to ignite each pixel according to the gray distribution of each pixel and the neighborhood of the pixel by the network so as to obtain a segmented binary image and extract a candidate target.
6. The adaptive target detection method based on SPCNN of claim 5, wherein: the specific implementation method of the step 5 is as follows,
step 5.1: calculating Hash values of output results of two adjacent iterations, namely H (i) and H (j), (i ═ 1,2, …, N, j ═ i + 1); calculating hamming distance between h (i) and h (j), which is denoted as hds (N), (N ═ 2, …, N);
step 5.2: the smaller Hds (n) is, the higher the similarity of the results of two adjacent iterations is, and then in the previous n iterations, the minimum value of Hds (n) is searched;
step 5.3: calculating the gray level average value of the segmentation result in each iteration, and recording the gray level average value as mean (N), (N is 1,2, …, N), wherein in order to avoid finding out that Hds (N) is a local minimum value, an additional condition mean (N) > preset threshold value needs to be added; when the two conditions are met, the optimal output result is obtained, and the judgment iteration is finished at the same time, so that the automatic selection of the optimal output result is completed, and the efficiency and the intellectualization of the target detection method are improved.
7. The adaptive target detection method based on SPCNN of claim 6, wherein: in step 1, Otsu's method is used to find the highest gray value in the background.
8. The adaptive target detection method based on SPCNN of claim 6, wherein: the preset threshold value in the step 5.3 is required to ensure that the detected target is meaningful, and is selected from 0.0095-0.025.
9. The adaptive target detection method based on SPCNN of claim 8, wherein: the preset threshold in step 5.3 is 0.01.
CN201910521356.2A 2019-06-17 2019-06-17 Adaptive target detection method based on SPCNN Active CN110276776B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910521356.2A CN110276776B (en) 2019-06-17 2019-06-17 Adaptive target detection method based on SPCNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910521356.2A CN110276776B (en) 2019-06-17 2019-06-17 Adaptive target detection method based on SPCNN

Publications (2)

Publication Number Publication Date
CN110276776A CN110276776A (en) 2019-09-24
CN110276776B true CN110276776B (en) 2021-06-15

Family

ID=67960848

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910521356.2A Active CN110276776B (en) 2019-06-17 2019-06-17 Adaptive target detection method based on SPCNN

Country Status (1)

Country Link
CN (1) CN110276776B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110969590B (en) * 2019-12-10 2022-05-27 兰州交通大学 Image enhancement algorithm based on CA-SPCNN
CN110889876B (en) * 2019-12-10 2022-05-03 兰州交通大学 Color image quantization method based on CA-SPCNN algorithm
CN111753853B (en) * 2020-07-08 2024-02-09 海南热带海洋学院 MPCNN-FAST sonar image feature point detection method
CN115861359B (en) * 2022-12-16 2023-07-21 兰州交通大学 Self-adaptive segmentation and extraction method for water surface floating garbage image
CN116703951B (en) * 2023-08-09 2023-10-20 成都理工大学 Image segmentation method based on random coupling neural network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106709497A (en) * 2016-11-16 2017-05-24 北京理工大学 PCNN-based infrared motion weak target detection method
CN107292883A (en) * 2017-08-02 2017-10-24 国网电力科学研究院武汉南瑞有限责任公司 A kind of PCNN power failure method for detecting area based on local feature

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106709497A (en) * 2016-11-16 2017-05-24 北京理工大学 PCNN-based infrared motion weak target detection method
CN107292883A (en) * 2017-08-02 2017-10-24 国网电力科学研究院武汉南瑞有限责任公司 A kind of PCNN power failure method for detecting area based on local feature

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
"Color image enhancement based on HVS and PCNN";ZHANG YUDONG等;《Information Science》;20101030;第53卷(第10期);全文 *
"一种参数自适应的简化PCNN图像分割方法";周东国等;《自动化学报》;20140630;第40卷(第6期);全文 *
"基于PCNN和改进邻域判决的红外弱小目标检测算法";李云等;《兵器装备工程学报》;20180131;第39卷(第1期);全文 *

Also Published As

Publication number Publication date
CN110276776A (en) 2019-09-24

Similar Documents

Publication Publication Date Title
CN110276776B (en) Adaptive target detection method based on SPCNN
CN112614077B (en) Unsupervised low-illumination image enhancement method based on generation countermeasure network
CN111369597B (en) Particle filter target tracking method based on multi-feature fusion
CN105374026B (en) A kind of detection method of marine infrared small target suitable for coast defence monitoring
CN111753881A (en) Defense method for quantitatively identifying anti-attack based on concept sensitivity
CN115063329A (en) Visible light and infrared image fusion enhancement method and system under low-illumination environment
CN111401246A (en) Smoke concentration detection method, device, equipment and storage medium
CN107403222A (en) A kind of motion tracking method based on auxiliary more new model and validity check
Hanmadlu et al. A novel optimal fuzzy color image enhancement using particle swarm optimization
CN115641348A (en) Method for determining pupil edge of eye based on user-defined area factor
Nalla et al. Image dehazing for object recognition using faster RCNN
CN107516083A (en) A kind of remote facial image Enhancement Method towards identification
CN108986083B (en) SAR image change detection method based on threshold optimization
CN112232440B (en) Method for realizing information memory and distinction of impulse neural network by using specific neuron groups
CN112305560B (en) Single photon laser radar rapid imaging method based on head photon group
Widyantara et al. Gamma correction-based image enhancement and canny edge detection for shoreline extraction from coastal imagery
CN111627030A (en) Rapid and efficient sea-land accurate segmentation method for visible light remote sensing image
CN113066077B (en) Flame detection method and device
Zheng et al. Image segmentation method based on spiking neural network with adaptive synaptic weights
Xiao et al. Automatic image segmentation algorithm based on PCNN and fuzzy mutual information
CN112541859B (en) Illumination self-adaptive face image enhancement method
Li et al. Image object detection algorithm based on improved Gaussian mixture model
CN116935496B (en) Electronic cigarette smoke visual detection method
CN113920159B (en) Infrared air small and medium target tracking method based on full convolution twin network
CN112270220B (en) Sewing gesture recognition method based on deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant