CN103268494A - Parasite egg recognition method based on sparse representation - Google Patents

Abstract

The invention belongs to the technical field of image recognition, and specifically relates to a method for identifying parasite eggs based on sparse representation, including: establishing an initial dictionary; using the K-SVD algorithm to learn the dictionary; processing an input image; calculating and reconstructing an error matrix; obtaining candidates Image blocks; steps such as identifying candidate image blocks. The invention introduces a classification algorithm based on sparse representation, which enhances the robustness of the entire parasite egg recognition algorithm to various interference factors; introduces the Batch-OMP algorithm for large-scale sparse representation process, and improves the recognition efficiency; introduces The method of directly building a dictionary with samples after Gaussian pyramid dimension reduction avoids the step of extracting egg target features, making the recognition process easier; introduces the method of establishing an error matrix and finding its local minimum, avoiding the Different image blocks containing the same object are obtained in the preliminary localization process.

Description

Parasite egg recognition method based on sparse representation

technical field

The invention belongs to the technical field of image recognition, and in particular relates to a method for recognizing parasite eggs based on sparse representation.

Background technique

The key to the automatic identification of parasite eggs based on computer image processing and medical microscopy technology is to design a fast and effective image recognition algorithm. In the past, the automatic identification method of parasite eggs based on images mainly relied on the separation of egg targets first. , and then extract various features of the target, and finally combine a classifier to complete the recognition. Give two examples of methods more relevant to the present invention: (1) Derya Avci et al. combined 7 invariant moments and support vector machines of Hu in 2009 in the document "An expert diagnosis system for classification of human parasite eggs based on multi-class SVM "Recognize 16 kinds of human parasite eggs. Although a high recognition rate is obtained, it can only be achieved under the premise that the image is relatively ideal, and the situation when there are many interference factors is not considered; (2) Chinese patent CN201110022426. 3 A method combined with the edge histogram of parasite eggs is proposed to recognize the shape of human parasite eggs, which overcomes the influence of weak boundaries and improves the reliability of recognition. However, for parasites with similar shapes Shape recognition is still lacking. Judging from the existing methods, there are many types of features. In addition to the features described in the above methods, they also include color, shape, size, texture, etc. The quality of feature selection largely determines the final recognition rate. And the step of extracting features is also difficult to complete accurately. There are also a variety of classifiers, including Bayesian classifiers, linear discriminant analysis, support vector machines, neural networks, minimum distance, etc., because these classifiers are sensitive to features, so which feature to choose is optimal for the classifier is often difficult to determine, and these classifiers are less robust to interference factors such as noise, occlusion, and impurities.

The application of classification algorithms based on sparse representation is far from being developed. Such a basic algorithm framework needs to be transformed and expanded in combination with other technologies and techniques in different application occasions, especially in data dimensionality reduction, selection of sparse representation algorithms, Dictionary learning needs to be determined according to specific needs. Based on the above analysis, it is applied to the identification of parasite eggs for the first time to realize the identification of single or multiple types of parasite eggs.

Contents of the invention

The purpose of the present invention is to overcome the defect that the previous parasite egg identification method is sensitive to features and various interference factors, and combines the Batch-OMP algorithm suitable for large-scale sparse representation and the K-SVD dictionary learning algorithm to propose a method based on sparse The parasite egg identification method expressed in the paper can meet the application requirements of the actual parasite egg automatic identification system in terms of recognition rate and identification efficiency.

In order to achieve the above-mentioned purpose of the invention, the present invention adopts the following technical solutions: a method for identifying parasite eggs based on sparse representation, comprising the following steps:

(1) Establish an initial dictionary: establish an initial single-class dictionary for single-class recognition, and establish an initial joint dictionary for multi-class recognition;

(2) Dictionary learning: use the K-SVD algorithm to learn the dictionary, single-class recognition to obtain a single-class representation dictionary, multi-class recognition to obtain a joint representation dictionary and a joint discrimination dictionary;

(3) Processing the input image: Pyramid compression is performed on the input image, and the compressed image is divided into blocks by means of a sliding window, and the step size can be selected as one or more pixels; all image blocks are sparsely represented, and a single-class recognition dictionary A single-class representation dictionary is used, and a multi-class recognition dictionary uses a joint representation dictionary;

(4) Calculate the reconstruction error matrix;

(5) Acquire candidate image blocks: use the reconstruction error matrix obtained in step (4) to find all its local minimum values, and select the image blocks corresponding to the smallest k values as candidate targets;

(6) Recognition of candidate image blocks: For single-class recognition, the threshold is used to distinguish candidate image blocks, and the recognition is completed; for multi-class recognition, the candidate image blocks are sparsely represented, and the joint discriminant dictionary is used to calculate the reconstruction error of the sub-dictionary, using The threshold method is used to discriminate and classify the candidate image blocks, and the recognition is completed.

In step (1), the steps to create an initial dictionary are as follows:

(1) Select a number of representative parasite egg image samples c·n with less impurities, where c is an integer ≥ 1, representing the number of classes, and n represents the number of samples of each class;

(2) Gaussian pyramid is used to compress c n images to obtain image samples after dimensionality reduction;

(3) With d degrees as the interval, rotate each image obtained in the previous step for one week to obtain 360/d image samples (including the original image), so the total number of samples is N=360 c n/d;

(4) "stretch" each two-dimensional image data obtained in the previous step into a one-dimensional vector, and then standardize each vector so that each vector satisfies the l ₂ -norm of 1;

(5) Use all the standardized vectors obtained in the previous step as the atoms of the dictionary to obtain the initial dictionary. If c=1, the initial single-class dictionary for single-class recognition is obtained. If c>1, the multi-class dictionary is obtained. The identified initial joint dictionary, which contains c sub-dictionaries.

In step (2), the K-SVD algorithm is used to learn the dictionary, which is divided into three situations:

(1) For the identification of single-type parasite eggs, the K-SVD algorithm is used to learn the initial single-type dictionary to obtain a single-type representation dictionary, which is used for preliminary positioning and classification at the same time. The volume of the dictionary is based on the dimension of the atomic vector depends;

(2) For the identification of multiple types of parasite eggs, use the K-SVD algorithm to learn the entire initial joint dictionary to obtain a joint representation dictionary, which is used for preliminary positioning;

(3) For the identification of multiple types of parasite eggs, use the K-SVD algorithm to learn each initial sub-dictionary, and then combine all the learned sub-dictionaries to obtain a joint discriminant dictionary. This dictionary is used for classification and its volume is much larger than The union represents the volume of the dictionary.

In step (3), the sparse representation of all image blocks is a large-scale sparse representation, that is, the Batch-OMP algorithm is used to solve the formula (1-1):

min||x-Dθ|| ₂ st||θ|| ₀ ≤ T (1-1)

where x is the input signal, D is the single-class representation dictionary or joint representation dictionary obtained in step (2), θ is the coefficient, and T is the sparsity condition.

In step (4), calculate the reconstruction error matrix, the steps are as follows:

(1) Use the formula (1-2) to calculate the reconstruction error, and obtain the reconstruction error [e ₁ , e ₂ ,...e _L ] of all image blocks, where L is the number of image blocks;

$\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D\&theta;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

(2) Arrange [e ₁ , e ₂ ,...e _L ] into a two-dimensional matrix according to the order of the image blocks obtained in step (3), and the obtained reconstruction error matrix is obtained.

In step (6), the sub-dictionary reconstruction error is calculated according to the formula (1-3).

$\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Where D _i is a sub-dictionary of the joint dictionary D=[D ₁ , D ₂ , . . . D _c ], where i=1, 2, . . . , c, where c is the number of classes.

The method for identifying parasite eggs based on sparse representation of the present invention introduces a classification algorithm based on sparse representation, which enhances the robustness of the entire parasite egg identification algorithm to various interference factors; introduces the Batch-OMP algorithm for The large-scale sparse representation process improves the recognition efficiency; introduces the method of directly building a dictionary with samples after Gaussian pyramid dimensionality reduction, avoids the step of extracting egg target features, and makes the recognition process easier; introduces the establishment error The method of finding the local minimum value of the matrix avoids getting different image blocks containing the same target in the preliminary positioning process.

Description of drawings

Fig. 1 is a flow chart of the method for identifying parasite eggs based on sparse representation in the present invention.

Figure 2 Selected several types of parasite egg samples.

Figure 3. Initial one-class dictionary.

Figure 4. Initial joint dictionary.

Figure 5 represents a dictionary with a single class.

Figure 6 jointly represents dictionaries.

Fig. 7 Joint discriminant dictionary.

Figure 8 Input image.

Figure 9 Reconstruction error vector.

Figure 10 Reconstruction error matrix.

Detailed ways

The present invention will be further explained below in conjunction with specific examples.

As shown in Figure 1, it is the process flow of the method for identifying parasite eggs based on sparse representation of the present invention, including the following steps:

(1) Establish an initial dictionary: establish an initial single-class dictionary for single-class recognition, and establish an initial joint dictionary for multi-class recognition. Step 1: Select a number of representative parasite egg image samples c·n with less impurities, where c≥1 represents the number of classes, and n represents the number of samples of each class, as shown in Figure 2. Step 2: Compress c·n images with Gaussian pyramid to obtain dimensionally reduced image samples. Step 3: With d degrees as the interval, rotate each image obtained in the previous step for one week to obtain 360/d image samples (including the original image), so the total number of samples is N=360 c n/d, which can be taken d=5. Step 4: "stretch" each two-dimensional image data obtained in the previous step into a one-dimensional vector, and then standardize each vector so that each vector satisfies the l ₂ -norm of 1. Step 5: Use all the standardized vectors obtained in the previous step as the atoms of the dictionary to obtain the initial dictionary. If c=1, the initial single-class dictionary is obtained, as shown in Figure 3. If c>1, the obtained is the initial joint dictionary of multiple classes, including c sub-dictionaries, as shown in Figure 4.

(2) Dictionary learning: The K-SVD algorithm is used to learn the dictionary, a single-class representation dictionary is obtained for single-class recognition, and a joint representation dictionary and a joint discriminant dictionary are obtained for multi-class recognition. There are three cases. The first one: for the identification of single-type parasite eggs, the K-SVD algorithm is used to learn the initial single-type dictionary to obtain a single-type representation dictionary, as shown in Figure 5, the dictionary is used for preliminary positioning and classification at the same time, the dictionary The volume depends on the dimensionality of the atomic vectors. The second method: for the identification of multi-type parasite eggs, the K-SVD algorithm is used to learn the entire initial joint dictionary to obtain a joint representation dictionary, as shown in Figure 6, which is used for preliminary positioning. The third method: for the identification of multi-type parasite eggs, use the K-SVD algorithm to learn each initial sub-dictionary, and then combine all the learned sub-dictionaries to obtain a joint discriminant dictionary, as shown in Figure 7, the dictionary is used For classification, its volume is much larger than that of the joint representation dictionary.

(3) Processing the input image: Pyramid compression is performed on the input image, and the compressed image is divided into blocks by means of a sliding window, and the step size can be selected as one or more pixels; all image blocks are sparsely represented, and a single-class recognition dictionary A single-class representation dictionary is used, and a joint representation dictionary is used for multi-class recognition dictionaries.

In step (3), sparse representation of all image patches involves large-scale sparse representation, and formula (1-1) is solved using Batch-OMP algorithm.

min||x-Dθ|| ₂ st||θ|| ₀ ≤ T (1-1)

where x is the input signal, D is the representation dictionary or joint representation dictionary obtained in step (3), θ is the coefficient, and T is the sparsity condition.

(4) Calculate the reconstruction error matrix. Step 1: Use the formula (1-2) to calculate the reconstruction error, and obtain the reconstruction error [e ₁ , e ₂ ,...e _L ] of all image blocks, as shown in Figure 8 (input image) and Figure 9 (reconstruction error vector ) shown.

$\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D\&theta;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Step 2: Arrange [e ₁ , e ₂ ,...e _L ] into a two-dimensional matrix according to the position of the image block in the original image, and the obtained reconstruction error matrix is shown in Figure 10.

(5) Acquire candidate image blocks: Use the reconstruction error matrix obtained in step (4) to find all its local minimum values, and select the image blocks corresponding to the smallest k values as candidate targets, as shown in Figure 10. distribution of minimum points.

(6) Recognition of candidate image blocks: For single-class recognition, the threshold is used to distinguish candidate image blocks, and the recognition is completed; for multi-class recognition, the candidate image blocks are sparsely represented, and the joint discriminant dictionary is used to calculate the reconstruction error of the sub-dictionary, using The threshold method is used to discriminate and classify the candidate objects. The candidate image blocks whose reconstruction errors are greater than the threshold are judged to contain no known objects. The corresponding class is identified.

In this step, the sub-dictionary reconstruction error is calculated according to formula (1-3).

$\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Where D _i is a sub-dictionary of the joint dictionary D=[D ₁ , D ₂ , . . . D _c ], where i=1, 2, . . . , c, where c is the number of classes.

Claims (6)

1. the parasite egg recognition methods based on rarefaction representation is characterized in that, may further comprise the steps:

(1) set up initial dictionary: initial single category dictionary is set up in single class identification, and initial associating dictionary is set up in multiclass identification;

(2) dictionary study: use the K-SVD algorithm that dictionary is learnt, single class identification obtains single class and represents dictionary, and multiclass identification obtains uniting the expression dictionary and unites the differentiation dictionary;

(3) handle input picture: input picture is carried out pyramidal compression, with the mode of moving window compressed image is carried out piecemeal, step-length can be chosen as one or more pixels; The all images piece is carried out rarefaction representation, and the dictionary of single class identification adopts single class to represent dictionary, and the dictionary of multiclass identification adopts associating expression dictionary;

(4) calculate the reconstruction error matrix;

(5) obtain the candidate image piece: utilize the reconstruction error matrix that obtains in the step (4), seek its all local minimum, choose k the wherein minimum corresponding image block of value as candidate target;

(6) identification candidate image piece: for single class identification situation, differentiate the candidate image piece with threshold value, identification is finished; For multiclass identification situation, the candidate image piece is carried out rarefaction representation, use associating differentiation dictionary, calculate sub-dictionary reconstruction error, use threshold mode that the candidate image piece is differentiated and classification, identification is finished.

2. the parasite egg recognition methods based on rarefaction representation according to claim 1 is characterized in that: in the step (1), the initial dictionary step of described foundation is as follows:

(1) select the less and representative parasite egg image pattern cn of some impurity, wherein c is 〉=1 integer, represents the class number, and n represents the sample number of each class;

(2) adopt gaussian pyramid that cn width of cloth image is compressed, obtain the image pattern behind the dimensionality reduction;

(3) be spacing with the d degree, the every width of cloth image that obtains in the previous step is rotated a circle obtains 360/d image pattern (comprising former figure), so total sample number is N=360cn/d;

(4) each two-dimensional image data " elongation " that previous step is obtained is one-dimensional vector, again each vector is carried out standardization, makes each vector satisfy l ₂-norm is 1;

(5) all standardized vectors that previous step is obtained obtain initial dictionary as the atom of dictionary, if c=1, what then obtain is initial single category dictionary of single class identification, if c＞1, what then obtain is the initial associating dictionary of multiclass identification, comprises c sub-dictionary.

3. the parasite egg recognition methods based on rarefaction representation according to claim 1 is characterized in that: in the step (2), described use K-SVD algorithm is learnt dictionary, is divided into three kinds of situations:

(1) at single class parasite egg identification, with the K-SVD algorithm initial single category dictionary is learnt, obtained single class and represent dictionary, this dictionary is used for Primary Location and classification simultaneously, and the volume of dictionary is decided according to the dimension of former subvector;

(2) at the identification of multiclass parasite egg, with the K-SVD algorithm whole initial associating dictionary is learnt, obtained associating expression dictionary, this dictionary is used for Primary Location;

(3) at multiclass parasite egg identification, with the K-SVD algorithm each initial sub-dictionary is learnt, the sub-dictionary after all study is united obtain associating differentiation dictionary again, this dictionary is used for classifying, and its volume is much larger than uniting the volume of representing dictionary.

4. the parasite egg recognition methods based on rarefaction representation according to claim 1 is characterized in that: in the step (3), described all images piece is carried out rarefaction representation is extensive rarefaction representation, namely uses Batch-OMP algorithm solution formula (1-1)

min||x-Dθ|| ₂s.t.||θ|| ₀≤T （1-1）

Wherein x is input signal, and D is that the single class that obtains in the step (2) is represented dictionary or united the expression dictionary, and θ is coefficient, and T is sparse property condition.

5. the parasite egg recognition methods based on rarefaction representation according to claim 1 is characterized in that: in the step (4), and described calculating reconstruction error matrix, step is as follows:

(1) utilizes formula (1-2) to calculate reconstruction error, obtain the reconstruction error [e of all images piece ₁, e ₂... e _L], wherein L is the image block number;

$e={\left|\right|x-\mathrm{D\&theta;}\left|\right|}_2^2---(1-2)$

(2) order of the image block that obtains according to step (3) is with [e ₁, e ₂... e _L] be arranged in a two-dimensional matrix according to the order of sequence, what obtain is the reconstruction error matrix.

6. the parasite egg recognition methods based on rarefaction representation according to claim 1 is characterized in that: in the step (6), the sub-dictionary reconstruction error of described calculating calculates according to formula (1-3)

$e={\left|\right|x-D_i\mathrm{\&theta;}\left|\right|}_2^2---(1-3)$

D wherein _iBe associating dictionary D=[D ₁, D ₂... D _c] sub-dictionary, i=1 wherein, 2 ..., c, c are the class number.

Images