CN111652317B - Super-parameter image segmentation method based on Bayes deep learning - Google Patents

Abstract

The invention discloses a super-parameter image segmentation method based on Bayes deep learning, and aims to solve the technical problems of large calculation amount and low precision of super-parameter extraction in the existing image segmentation. The method comprises the steps of selecting an image training set, performing a Gaussian process on image information, preprocessing a data set by adopting an L2 regular operator to obtain image contour edge features, constructing a target feature edge feature recognition training set, classifying the data set according to a Bayesian theorem, setting an image target edge segmentation label based on semantic recognition, extracting the target edge feature data set by further adopting the Gaussian process, and calculating a target set edge feature Gaussian super-parameter set. The invention has the beneficial effects that: and the efficiency and the accuracy of target identification are improved.

Description

Hyperparameter Image Segmentation Method Based on Bayesian Deep Learning

technical field

The invention relates to the technical field of machine learning, in particular to a hyperparameter image segmentation method based on Bayesian deep learning.

Background technique

In computer vision, image segmentation refers to the task of assigning a label to each pixel in an image, which can also be viewed as pixel classification. Unlike object detection using rectangular candidate boxes, image segmentation needs to be accurate to pixel-level locations, so it plays a very important role in tasks such as medical analysis, satellite image object detection, iris recognition, and self-driving cars.

Humans recognize targets more by relying on experience to distinguish targets, while deep learning is to build a convolutional neural network, rely on training to extract target features, and then identify targets. The recognition results of traditional target recognition methods are often some pre-defined objects of a certain category, such as faces, vehicles, etc., and the content contained in an image is far more than some independent objects. It also contains multiple Objects and their attributes, spatial relationships, logical relationships and other information, these information cannot be described with only some class labels, but need to be described in natural language. It is difficult for any mathematical model to satisfy all target recognition, so many conditional classification and recognition are formed, resulting in low efficiency of deep learning cross-domain fusion recognition.

Pixel-level image segmentation is a research hotspot in the field of artificial intelligence. It is a mathematical modeling problem involving multiple disciplines such as image processing, pattern recognition, visual perception and psychological cognition. It is very easy for human beings to rely on experience to identify targets in the long-term evolution and learning process, but relying on machines to automatically identify targets from complex backgrounds requires complex mathematical modeling to achieve, so the identification model and hyperparameter optimization are selected appears to be important. The selection of hyperparameters in deep learning is difficult and irregular, and there are unpredictable effects between different hyperparameters. Its debugging is very time-consuming, and the evaluation of each hyperparameter combination requires a large number of iterative calculations. For such problems, some classic optimization algorithms such as particle swarm optimization algorithm, simulated annealing algorithm, local search algorithm, etc. are no longer applicable. Some researchers have proposed the method of using a proxy model to reduce the evaluation cost of this type of problem by simulating the estimated value of the objective function. However, no matter using the Monte Carlo algorithm or the adaptive simulation algorithm proposed in the field of reinforcement learning, the learning process is always time-consuming, and it is limited to a certain condition or field, and the accuracy is difficult to guarantee, and it is difficult to achieve cross- world fusion.

Contents of the invention

The present invention provides a hyperparameter image segmentation method based on Bayesian deep learning to solve the existing technical problems of large amount of calculation and low precision in the extraction of hyperparameters in image segmentation.

In order to solve the problems of the technologies described above, the present invention adopts the following technical solutions:

Design a hyperparameter image segmentation method based on Bayesian deep learning, including:

Step 1: Data preprocessing, normalize the data elements in the image, and generate an image segmentation data set;

Step 2: Use Gaussian mask to extract target edge features on the image;

Step 3: Using Bayesian estimation, extract the bounding box and target mask of the image through the target edge feature;

Step 4: Put the bounding box and the target mask into the feature dictionary for comparison, and then the category of each target in the image can be obtained;

The construction method of the feature dictionary is as follows:

(1) set up a training set and a test set of images;

(2) Perform the operations in the above steps 1-3 for each image in the training set to obtain its bounding box and target mask;

(3) Collect the bounding boxes and target masks in (2) to obtain a feature dictionary composed of them;

(4) Input the test set into the feature dictionary in (3), and check the accuracy rate of the obtained feature dictionary. If the accuracy rate does not meet the requirements, adjust the parameters of the model and retrain until the accuracy rate of the feature dictionary meets the requirements.

Further, in step 1, the image segmentation data set contains N target segmentation class attributes and M data attributes of each target class. When the probability of N class attributes and M data attributes is the largest, Bayesian Classification matcher, select the target and segment the image; the software used is python, and the framework adopts tensorflow.

Further, in step 2, the specific steps of extracting target edge features are:

Step 1: Suppose the edge probability of the image pixel f(x,y) satisfies the Gaussian distribution, then its two-dimensional Gaussian function is:

The second step: Find the gradient function for the x and y directions of the above image:

Step 3: Convolve the image dataset:

Step 4: Calculate the image target edge probability density distribution, that is, the target edge features:

Further, in step 2, before extracting the target category label, calculate the prior probability of the target appearing:

Among them, C _i is any element in the target set of type C (C ₁ , C ₂ , C ₃ . . . C _n ), N _i represents the number of occurrences of the target, and N represents the total amount of the target set.

Further, in step 2, in the process of extracting the target category label, the conditional probability of the target appearance is calculated:

Among them, x _a represents the abscissa of the target point, and y _a represents the ordinate of the target point. P(x _a ), P(y _a ) represent the target edge feature probability.

Further, in step 3, the specific steps of extracting the bounding box and target mask of the image are:

(1) Obtain the target area of the image and the classification weight of each pixel in the area by learning the extracted target boundary features;

(2) After obtaining the target area of the image, combine the internal and external feature maps of each target area into two complete feature maps, and then perform image segmentation and image classification two branch data sets D1 and D2 simultaneously;

(3) In image segmentation, a Bayesian classifier is used to classify the inner and outer feature maps of the target region to distinguish the foreground and background in the image and generate a mask;

(4) In image classification, take the maximum value according to the pixel probability distribution in the two types of feature maps to obtain a new feature map, and then use the maximum likelihood estimation classifier to obtain the object category in the target area.

Further, in step 4, the method of comparing the bounding box and the target mask with the feature dictionary is as follows: first, use the L2 regular operator to calculate the similarity weights of the bounding box and the target mask and the feature dictionary respectively, and then the similarity Gaussian The process extracts the target edge feature data set, and after Bayesian classification matching, the semantic segmentation result can be obtained.

Furthermore, before the output of the semantic segmentation results, the edge Gaussian hyperparameter function is calculated first, and then the target matching degree is calculated according to the score. The better the hyperparameter set, the higher the accuracy score of the semantic segmentation.

Further, the training set of images includes the Open Images V4 detection set, which contains 1.9 million images and 15.4 million bounding boxes for 600 categories on the images.

Compared with the prior art, the beneficial technical effect of the present invention is:

The present invention mainly utilizes Bayesian formula, python language, tensorflow framework, according to the principle that after the Gaussian process of the edge feature of the image, the protrusion of the edge feature is steep, performs pixel preprocessing on the image, and performs a Gaussian process on the entire image to obtain the edge feature Data set, and then use the L2 criterion to preprocess the data set, and then according to the image target segmentation requirements based on semantic recognition, use the Bayesian estimation model for image target edge feature recognition, and combine deep learning to construct image target edge based on semantic recognition The feature data dictionary applies the trained model to the complex target recognition system, improving the efficiency and accuracy of target recognition.

Description of drawings

Fig. 1 is the flow chart of the hyperparameter image segmentation method based on Bayesian deep learning of the present invention.

Detailed ways

The specific implementation of the present invention will be described below in conjunction with the accompanying drawings and examples, but the following examples are only used to describe the present invention in detail, and do not limit the scope of the present invention in any way.

The programs involved or relied on in the following embodiments are conventional programs or simple programs in the technical field, and those skilled in the art can make conventional selections or adaptive adjustments according to specific application scenarios.

Embodiment 1: A hyperparameter image segmentation method based on Bayesian deep learning, see Fig. 1, its overall steps are: select image training set, carry out Gaussian process to image information, adopt L2 regular operator to carry out pre-processing to data set Processing, obtaining image contour edge features, and constructing target feature edge feature recognition training set, classifying the data set according to Bayesian theorem, and setting image target edge segmentation labels based on semantic recognition, and further using Gaussian process to extract target edges Feature data set, and calculate the target set edge feature Gauss hyperparameter set, calculate the target edge posterior probability according to the data set, and take the maximum posterior probability as the target image segmentation and recognition probability based on semantics, and then pass the Bayesian score Matching, if the score is higher than 90, the recognition is considered correct, otherwise, Gaussian and hyperparameters are adjusted by deep learning and 0.618 coefficients, and Gaussian process training is performed again until super-optimal parameters are obtained. The final result of the present invention can be achieved: input images and target labels into the model for image target segmentation, and the target can be separated from the corresponding background. That is, for a trained model, given an image and the target information to be queried, the corresponding target can be detected from the image.

The method of making the above feature dictionary is as follows:

(1) Data preprocessing

Regularize the data elements in the image to generate an image segmentation dataset. The data set contains N target segmentation class attributes and M data attributes representing each target class. When the probability of N class attributes and M data attribute probabilities are the largest, the model is used to segment the image according to the target edge contour. The specific implementation uses the Python language and is implemented under the artificial intelligence tensorflow framework.

(2) Extract target edge features

Mainly through the Gaussian transformation to extract the target super-edge parameter features, the main extracted features are as follows:

1. The gray level of the edge pixel of the target image jumps;

2. The pixel boundary between different materials, textures, colors and brightness of the target changes abruptly;

3. The target contour line and the background have different reflection characteristics, which will also cause a jump in the pixel value;

4. When the target is illuminated, it will form a shadow, which will also cause a jump in the gray value between pixels.

Calculating the above attributes includes: using a Gaussian mask edge feature extraction algorithm for the image f(x, y), which is characterized in that: according to the feature of the edge data jump of the image target and the feature with a large probability density change, an appropriate Gaussian mask is selected , based on the calculation of the maximum value in the neighborhood of the pixel point, the position with high probability density of the data set is obtained, which is the edge pixel point.

The specific process is as follows:

Step 1: Suppose the edge probability of the image pixel f(x,y) satisfies the Gaussian distribution, then the two-dimensional Gaussian function is:

The second step: Find the gradient function for the x and y directions:

Step 3: Convolve the image dataset:

Step 4: Calculate the image target edge probability density distribution, that is, the target edge features:

The above is the calculation method of the Gaussian edge hyperparameter feature extraction module. The essence is to calculate the Gaussian kernel function, and then use the Gaussian kernel function as a mask to convolve with the target pixel to extract the target edge feature. This process is also a hyperparameter estimation process.

After obtaining the image pixel edge features, the Gaussian process is used to calculate the feature distribution parameters, and then the L2 regularization operator is used to calculate the minimum (optimum) loss function of the Gaussian feature function. At the same time, a penalty function is added to the algorithm to prevent the model from overfitting. The L2 regularization operator is:

Among them, Loss is the loss function, E _in is the training sample error that does not contain the regularization item, and λ is the regularization parameter (penalty function). In order to make the model more optimized, the regular function is defined as follows:

That is, the sum of squares of all w (errors) does not exceed the parameter C (threshold), which can ensure that the training sample error E _in is minimized, and the loss function value is the smallest.

(3) Extract the bounding box and target mask of the image

The Bayesian estimation algorithm is mainly used. The Bayesian estimation model uses the prior probability recognition algorithm to recognize the objective world. Its basic idea is to estimate the maximum probability of the cognitive object through the prior probability and conditional probability on the basis of accumulating a large number of samples. The largest is the cognitive result. When the analysis sample is large enough to be close to the population, the probability of the event occurring in the sample will be close to the probability of the event occurring in the population, so the minimum error prediction can be realized. The core of Bayesian estimation is the selection of hyperparameters. In order to improve the local advantage in the process of image segmentation, a hyperparameter image data classification model based on Gaussian distribution is constructed, combined with deep learning, and the 0.618 estimation method (golden section method) is adopted. Select hyperparameters, effectively adjust the balance between hyperparameter errors and weights, and realize an efficient pixel segmentation method based on semantic understanding.

Bayesian classifier is a classification method based on statistical theory. For a sample set C={C ₁ C ₂ C ₃ ...... C _n } containing M class samples, the classifier first calculates N-dimensional features Vector X=[x ₁ x ₂ ...... x _n ] the maximum likelihood estimation of the labels belonging to each category, by sorting them and obtaining the maximum value to calculate the category label C _i to which x belongs, The Bayesian formula is as follows:

Among them, Pr( _ci |x) is the posterior probability, Pr(x| _ci ) is the conditional probability, Pr( _ci ) is the prior probability, P(x _a ), P(y _a ) represent the target edge feature probability . Then the classification problem boils down to the problem of finding the maximum value of x attribute class C _i :

C _i ＝argmaxPr(x|c _i )Pr(c _i ) (11)

Experiments have shown that the accuracy of Naive Bayes classifiers is much higher than that of other class classifiers.

According to the different gray levels of different images, there are generally obvious protruding edges at the boundary of the image, and this feature can be used to segment the image. The kernel operator is a filter with Gaussian hyperparameters, which has the characteristics of image denoising, smoothing and strengthening edge feature attributes. The calculation process is divided into four steps: the first step: smoothing the image with a Gaussian filter; the second step : Use the finite difference of the first-order partial derivative to calculate the magnitude and direction of the gradient; the third step: carry out non-maximum suppression on the gradient magnitude; the fourth step: detect and connect the edges with a double threshold algorithm.

The process of image segmentation is as follows:

(1) Obtain the target area of the image and the classification weight of each pixel in the area by learning the extracted target boundary features;

(2) After obtaining the target area of the image, combine the internal and external feature maps of each target area into two complete feature maps, and then perform image segmentation and image classification two branch data sets D1 and D2 simultaneously;

(3) In image segmentation, a Bayesian classifier is used to classify the inner and outer feature maps of the target region to distinguish the foreground and background in the image and generate a mask;

(4) In image classification, take the maximum value according to the pixel probability distribution in the two types of feature maps to obtain a new feature map, and then use the maximum likelihood estimation classifier to obtain the object category in the target area.

After the image segmentation is completed, an image target contour edge data set needs to be established, as follows: The task of image target edge segmentation includes the generation of target candidate sets, edge feature extraction of candidate targets, Bayesian classification of candidate targets, candidate target There are 5 basic sub-tasks, such as hyperparameter correction of the target algorithm, and construction of the candidate target edge feature dictionary. Alternative target datasets include the OpenImages V4 detection set, which contains 1.9 million images and 15.4 million bounding boxes on images for 600 categories. The pixel L2 regularization operator is used to preprocess the image data to form an image classification set and a classification image pixel subset, and a multidimensional Gaussian distribution probability model is used to calculate the target edge feature kernel mask, and the target edge feature data set is obtained through convolution. Bayesian evidence learning is used to obtain the prior probability of the target marginal feature classification set.

(4) Implementation process

It mainly includes three parts: (1) Image target pixel preprocessing: use L2 regularization algorithm, set threshold function to generate initial point set: X, Y=(x1, y1), (x2, y2),...(xt , yt); (2) Adopt Gaussian kernel model to construct data set D={(x1, y1)...(xt, yt)}; (3) Enter Bayesian maximum a posteriori probability estimation. The specific operation is:

A. Set labels according to the target category;

B. Statistical image target classification;

C. Calculate the prior probability of various targets: Pr(C _i );

D. According to the target classification, select all the data in the corresponding data set D, build a Gaussian process model, and extract the target edge point Xi, Yi set;

E. Calculate the Gaussian distribution hyperparameter function (μ _i , σ _i );

F. Further, use the acquisition function (μ _i , σ _i ) to calculate the next evaluation point xi, the value of i is 1~t, xi=argmaxu(x|D), and calculate the response yi;

G. Add new data points to the set D, D←D∪{xi,yi}, i←i+1;

H. Using the Bayesian posterior estimation formula to calculate the posterior probability

I. Calculate the posterior probability distribution: R=Max(Pr(C ₁ |X), Pr(C ₂ |X)...Pr(C _t X));

J. Test calculation results: Use the calculated posterior probability to match the target edge features. If the matching score is higher, the recognized month is closer to the actual target. The model sets score>90;

K. Correction, if the score is less than 90, use deep learning to correct the Gaussian hyperparameters μ _i and σ _i , and the correction coefficient is 0.618;

L. Putting target marginal probability parameters with a score greater than 90 into the data dictionary to form a hyperparameter data dictionary set.

(5) Training and testing

Divide the 1.9 million pictures of the above-mentioned Open Images V4 detection set into a training set and a test set, train the model through supervised learning, and input the test set to verify whether the trained model meets the requirements. So far, the feature dictionary has been established.

(6) Image detection

Input the image to be tested into the detection model, extract its bounding box and target mask, and then put the bounding box and target mask into the feature dictionary for comparison, and then the category of each target in the image can be obtained. The method is as follows: firstly, the L2 regular operator is used to calculate the similarity weights of the bounding box and the target mask and the feature dictionary respectively, and then the similarity Gaussian process extracts the target edge feature data set, and undergoes Bayesian classification matching, namely Semantic segmentation results can be obtained. Before the semantic segmentation results are output, first calculate the edge Gaussian hyperparameter function, and then calculate the target matching degree according to the score. The better the hyperparameter set, the higher the accuracy score of the semantic segmentation.

Through the above steps, the image segmentation method of hyperparameter optimization based on Bayesian deep learning can be completed, and the target edge segmentation hyperparameter dictionary can be obtained on the basis of image set training. With the hyperparameter dictionary, target recognition and segmentation can be realized, that is, after the system is trained, it can separate and recognize the target and the background through model calculation in the environment of input images and voice. This method can effectively solve the shortcomings of traditional deep learning optimization algorithms such as long time-consuming, large performance fluctuations, and large resource occupation. After the model is trained, it can be applied to a plug-in in smartphone image semantic recognition software.

The present invention has been described in detail above in conjunction with the accompanying drawings and embodiments. However, those skilled in the art can understand that each specific parameter in the above embodiments can also be changed without departing from the spirit of the present invention. A number of specific embodiments are formed, all of which are common variation scopes of the present invention, and will not be described in detail here.

Claims (4)

1. A super-parameter image segmentation method based on Bayes deep learning is characterized by comprising the following steps:

step 1: data preprocessing, namely regularizing data elements in an image to generate an image segmentation class data set;

step 2: extracting target edge features from the image by using Gaussian masks;

step 3: extracting a boundary box and a target mask of the image through the target edge characteristics by using Bayesian estimation;

step 4: the boundary frame and the target mask are put into a feature dictionary for comparison, so that the category of each target in the image can be obtained;

in step 2, the specific steps of extracting the edge feature of the target are as follows:

the first step: assuming that the edge probability of the image pixel f (x, y) satisfies the gaussian distribution, the two-dimensional gaussian function is:

and a second step of: gradient functions are obtained for the x and y directions of the image:

and a third step of: convolving the image dataset:

fourth step: calculating the probability density distribution of the image target edge, namely the target edge characteristics:

before the class label of the target is acquired in the step 3, the prior probability of the occurrence of the target needs to be calculated:

wherein C is _i For class C target set (C ₁ 、C ₂ 、C ₃ ...C _n ) Any one of the elements, N _i Representing the number of occurrences of the object, N representing the total amount of the object set,

in the step 3, in the process of obtaining the category label of the target, calculating the conditional probability of the occurrence of the target:

wherein x is _a Representing the abscissa, y of the target point _a Representing the ordinate of the target point, P (x _a )、P(y _a ) Representing the probability of the edge feature of the object,

in step 3, the specific steps of extracting the bounding box of the image and the target mask are as follows:

(1) Obtaining a target region of the image and a classification weight of each pixel in the region by learning the extracted target boundary characteristics;

(2) After the target areas of the images are obtained, the internal and external feature images of each target area are combined into two complete feature images, and then two branch data sets D1 and D2 of image segmentation and image classification are synchronously carried out;

(3) In image segmentation, classifying the internal and external feature images of the target region by using a Bayesian classifier to distinguish a foreground and a background in the image and generate a mask;

(4) In the image classification, the maximum value is taken according to the pixel probability distribution in the two types of feature images to obtain a new feature image, and then a maximum likelihood estimation classifier is used to obtain the class of the object in the target area;

in step 4, the method for constructing the feature dictionary is as follows:

(1) Establishing a training set and a testing set of images;

(2) Carrying out the operation in the step 1-3 on each image in the training set to obtain a boundary frame and a target mask;

(3) Collecting the bounding box and the target mask in the step (2), and obtaining a feature dictionary composed of the bounding box and the target mask;

(4) Inputting the test set into the feature dictionary in the step (3), checking the accuracy of the obtained feature dictionary, and if the accuracy does not meet the requirement, adjusting the parameters of the model to retrain until the accuracy of the feature dictionary meets the requirement.

2. The method for performing super-parametric image segmentation based on Bayesian deep learning according to claim 1, wherein in step 1, the image segmentation class data set comprises N target segmentation class attributes and M data attributes of each target class, and when the probability of the N class attributes is the maximum with the probability of the M data attributes, a Bayesian classification matcher is adopted to select the target and segment the image.

3. The method for segmenting a super-parametric image based on bayesian deep learning according to claim 1, wherein in step 4, the method for comparing the bounding box and the target mask with the feature dictionary is as follows: firstly, calculating similarity weights of a boundary frame and a target mask and the feature dictionary by using an L2 regular operator, then extracting a target edge feature data set in the similarity Gaussian process, and obtaining a semantic segmentation result through Bayesian classification matching.

4. A super-parametric image segmentation method based on bayesian deep learning according to claim 3, wherein before outputting the semantic segmentation result, an edge gaussian super-parametric function is calculated, then the target matching degree is calculated according to the score value, and the better the super-parametric set is, the higher the precision score of the semantic segmentation is.