A method for image denoising based on posterior probability

Abstract

The present invention discloses an image denoising method based on posterior probability. The method belongs to the field of computer vision, and the operation steps include: data collection, data labeling, Bayesian model training, classification results are poured into a dictionary learning sparse matrix for denoising, and image processing is performed on the denoised data. The image denoising method based on naive Bayesian posterior probability and dictionary learning sparse matrix described in the embodiment of the present invention, through the powerful classification function of the naive Bayesian model, finely classifies the noise data and sample label data in the image, and then provides the classification results to the dictionary learning sparse matrix for sparse expression, and achieves the effect of noise reduction and denoising through mathematical methods, improves the signal-to-noise ratio of medical image data, and provides technical support for computer clinical auxiliary diagnosis such as image segmentation and detection in the later stage, thereby improving the efficiency of automated clinical diagnosis in hospitals.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and relates to an application of dictionary learning and naive Bayesian classification related technologies in image processing. Specifically, it relates to a sparse matrix denoising method integrating Bayesian posterior probability classification and dictionary learning, which is used to provide preprocessing results for medical image processing and facilitate lesion detection and analysis of medical images. Background Art

[0002] The Bayesian classification method is a mathematical method based on the Bayesian principle. Based on statistical knowledge, sample data can be classified more accurately. According to the existing classification application scenarios, it can be concluded that the naive Bayesian method has a higher classification accuracy. This mathematical method combines the prior probability and posterior probability in probability statistics. Not only does it effectively avoid the interference of subjective factors on the classification results during the classification process, but it also has a high generalization ability, which reduces the overfitting of classification to a certain extent. The naive Bayesian method simplifies some processes of the Bayesian method and makes certain assumptions about the classified sample data, so that the sample data can be applied to the naive Bayesian method on the basis of mutual conditional independence. In this way, the fairness of the data is guaranteed during the sample data analysis process, and the naive Bayesian method reduces the complexity of Bayesian classification. At the same time, it can effectively improve the generalization ability of the Bayesian model in the classification process in real scenarios.

[0003] Dictionary learning is a feature extraction method, and its core idea is to use the least resources to express the most meaning. For example, in the past, human communication could express each other's emotions through words, but the emotions between each other had many results. The number of words was limited, but a limited number of words could be used to express infinite kinds of emotions. In this link, we can express infinite content with only the limited words stored in the dictionary, and the dictionary provides the basis for the appropriate arrangement of our emotional expression. Therefore, from a theoretical point of view, dictionary learning is to use limited features to express more core content in sample data. The dictionary that can extract the most essential features obtained in this way is our tool for dimensionality reduction of sample data. With this tool, we can filter out some unimportant dirty data or noise data, thereby reducing the overall analysis accuracy of these unfavorable factors on sample data.

[0004] Existing image denoising methods have the disadvantages of poor effect, complex process and low precision. In order to solve this problem, this paper uses the posterior probability of the naive Bayes method to classify the noise data and effective feature data in the medical image data, and obtains relatively accurate classification results, which are then used as the input data of the dictionary learning sparse matrix. The dictionary learning sparse matrix can remove the redundant noise data in the image and retain the label data in the image as much as possible. It can not only simplify the model to a certain extent and achieve the effect of dimensionality reduction, but also improve the signal-to-noise ratio of the image data, and provide a more accurate data basis for the later analysis and detection of medical images, thereby simplifying the process for doctors to grasp the lesions and helping to improve the level of medical automation. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide an image denoising method based on posterior probability, which utilizes the denoising function of the posterior probability of the naive Bayes method and the dictionary learning sparse matrix to improve the signal-to-noise ratio of image data.

[0006] The present invention provides the following technical solutions:

[0007] On the one hand, the present invention provides a denoising method for medical image data, which integrates a naive Bayes model and a dictionary learning sparse matrix, and has a relatively effective effect on improving the signal-to-noise ratio of image data.

[0008] Secondly, the method provided by the present invention has certain clinical medical significance, especially in the process of computer-aided diagnosis.

[0009] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention.

[0010] Specifically, the image denoising method based on posterior probability described in the present invention has the following specific operating steps:

[0011] (1.1) Data collection: label the collected data;

[0012] (1.2) Data labeling: Label data is labeled using the labelImg tool.

[0013] (1.3) Training of Bayesian model: Use Bayesian classification mathematical method to sub-classify the labeled data, solidify the trained parameters through the corresponding model file, and use it as the basis for later classification;

[0014] (1.4) The classification results are poured into the dictionary learning sparse matrix for noise reduction: The classification results are subjected to noise reduction by using the sparse expression function of the sparse matrix;

[0015] (1.5) Perform image processing on the denoised data;

[0016] In step (1.1), the acquired data is raw data scanned from CT and MR equipment;

[0017] In step (1.2), the data label annotation specifically refers to the annotation of label data noise data and sample label data;

[0018] After label calibration using the labelImg tool, the mapping relationship between category labels is stored in a yaml file;

[0019] In step (1.3), the model of the Bayesian classification mathematical method is as follows:

[0020]

[0021] Where i represents the sample number; f represents the sample set; p(C=c) represents the prior probability; p(F i =f i |C=c) represents independent probability distribution; n represents the number of samples; |F i represents the scaling factor; C represents the categorical variable;

[0022] In step (1.4), the dictionary learning sparse expression model is as follows:

[0023]

[0024] Where X represents the sparse expression result image; I represents the image output after Bayesian classification; s represents the number of patches the image is divided into; Represents the transposed matrix of the image path matrix; E s represents the extracted image path matrix; Z represents the noisy image data; D represents the sparse expression matrix; a s Represents the sparse expression signal data of the patch image.

[0025] Beneficial effect: Compared with the prior art, the image denoising method based on naive Bayes posterior probability and dictionary learning sparse matrix described in the embodiment of the present invention, through the powerful classification function of the naive Bayes model, finely classifies the noise data and sample label data in the image, and then provides the classification results to the dictionary learning sparse matrix for sparse expression, achieves the effect of noise reduction and denoising through mathematical methods, improves the signal-to-noise ratio of medical image data, provides technical support for later computer clinical assisted diagnosis such as image segmentation and detection, thereby improving the efficiency of hospital automated clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is an operation flow chart of the present invention;

[0027] Figure 2 It is a schematic diagram of the working example of the present invention in engineering application. DETAILED DESCRIPTION

[0028] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings, and the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0029] The invention discloses an image denoising method based on posterior probability, which uses the posterior probability classification of the naive Bayes model to classify the noise data and sample label data in the image data more accurately. The classification result can obtain relatively good cleaned image data. Then, based on these cleaned data, they are input into the dictionary learning sparse matrix, and the sparse expression function of the dictionary learning self-learning is used to reduce or remove the noise data in the image, so as to perform more detailed denoising on the basis of the classified image, and finally improve the signal-to-noise ratio of the image data, obtain high-quality preprocessed data, and facilitate more accurate image analysis in the later stage.

[0030] Furthermore, given medical image data (such as CTA, MRI, PET, etc.) can be used as input data for the Bayesian classification model. The data information is encapsulated in matrix form, using Python as the basic development language and the Python open source data package numpy as a tool for processing image data. This can save development time and reduce development difficulty, thus saving actual development costs.

[0031] Specifically, the image denoising method based on posterior probability described in the present invention has the following specific operating steps:

[0032] (1.1) Data collection: label the collected data;

[0033] The collected data can be directly scanned from CT, MR and other equipment, and the data obtained here is the original data;

[0034] (1.2) Data labeling: Label data is labeled using the labelImg tool. The labelImg tool is an open source dataset labeling tool. By manually labeling the target points in the image, a mapping relationship between the image and the label data can be generated and stored in a yaml file. Later, the model training can directly obtain data by reading the yaml file.

[0035] This part of the data requires the participation of professional doctors. The label data, noise data and sample label data are labeled using the labelImg tool to calibrate the labels, and the mapping relationship between the category labels is stored in the yaml file;

[0036] (1.3) Training of the Bayesian model: In this process, the Bayesian classification mathematical method is used to sub-classify the labeled data, and the trained parameters are solidified through the corresponding model file, which is used as the basis for later classification;

[0037] The model of the Bayesian classification mathematical method is as follows:

[0038]

[0039] Where i represents the sample number; f represents the sample set; p(C=c) represents the prior probability p(F i =f i |C=c) represents independent probability distribution; n represents the number of samples; |F i represents the scaling factor; C represents the categorical variable.

[0040] (1.4) The classification results are poured into the dictionary learning sparse matrix for noise reduction: In this step, the noise in the original data and the sample label data have been classified more accurately, so here we only need to use the sparse expression function of the sparse matrix to perform a noise reduction process on the noise data in the image;

[0041] The dictionary learning sparse expression model is as follows:

[0042]

[0043] Where X represents the sparse expression result image; I represents the image output after Bayesian classification; s represents the number of patches the image is divided into; Represents the transposed matrix of the image path matrix; E s represents the extracted image path matrix; Z represents the noisy image data; D represents the sparse expression matrix; α s Represents the sparse expression signal data of the patch image.

[0044] (1.5) Perform image processing on the denoised data; for example, perform image recognition and segmentation based on the denoised data;

[0045] Specifically, the data after noise reduction processing can be used for further analysis, such as image segmentation, lesion detection, etc.

[0046] Example 1

[0047] Embodiment 1 of the present invention provides an image denoising method based on naive Bayes posterior probability and dictionary learning sparse matrix, which utilizes the denoising effect of naive Bayes precise posterior probability and dictionary learning sparse matrix to classify the given medical image data into noise and label data, and obtain the cleaned preprocessing classification result; then the classification result is used as the input result of the dictionary learning sparse matrix, and the expression function of the dictionary learning sparse matrix is ​​utilized to perform denoising and noise reduction processing on the noise data, thereby improving the signal-to-noise ratio of the medical image data, providing more accurate and high-quality medical image data for the observation and analysis of later medical clinical data, and helping to improve the level of clinical diagnosis of doctors.

[0048] The image denoising method based on naive Bayesian posterior probability and dictionary learning sparse matrix described in the embodiment of the present invention, through the powerful classification function of the naive Bayesian model, finely classifies the noise data and sample label data in the image, and then provides the classification result to the dictionary learning sparse matrix for sparse expression, and achieves the effect of noise reduction and denoising through mathematical methods, thereby improving the signal-to-noise ratio of medical image data, as described in the following formula:

[0049] C=Rlb(1+S / N)

[0050] Among them, C represents the maximum information rate that the channel can transmit, which is referred to as the channel capacity; B represents the bandwidth of the channel, which is the bandwidth of the transmission channel; S / N represents the signal-to-noise ratio of the received signal, and the core is to obtain higher quality image data by improving S / N; 1b represents the logarithm with base 2;

[0051] It provides technical support for later computer-aided clinical diagnosis such as image segmentation and detection, thereby improving the efficiency of automated clinical diagnosis in hospitals; the invention has certain clinical application value and engineering practical value.

Claims

1. A method for image denoising based on posterior probability, characterized in that: The posterior probability classification of the naive Bayes model is used to classify the noise data and sample label data in the image data. The classification result can obtain the cleaned image data. Then, based on these cleaned data, they are input into the dictionary learning sparse matrix, and the sparse expression function of the dictionary learning self-learning is used to reduce or remove the noise data in the image, so as to perform noise reduction on the basis of the classified image, and finally improve the signal-to-noise ratio of the image data, obtain preprocessed data, and facilitate the subsequent image analysis. The specific steps are as follows: (1.1) Data collection: label the collected data; The collected data is directly scanned from CT and MR equipment; (1.2) Data labeling: Label data is labeled using the labelImg tool; the labelImg tool is an open source dataset annotation tool that manually annotates target points in an image to generate a mapping relationship between the image and the label data. The mapping relationship is stored in a yaml file, and the model training is then performed directly by reading the yaml file to obtain data. The label data noise data and sample label data are labeled using the labelImg tool, and the mapping relationship between category labels is stored in a yaml file; (1.3) Training of the Bayesian model: In this process, the Bayesian classification mathematical method is used to sub-classify the labeled data, and the trained parameters are solidified through the corresponding model file, which is used as the basis for later classification; The model of the Bayesian classification mathematical method is as follows: Where i represents the sample number; f represents the sample set; p(C=c) represents the prior probability; p(F i =f i |C=c) represents independent probability distribution; n represents the number of samples; F i represents the scaling factor; C represents the categorical variable; (1.4) The classification results are poured into the dictionary learning sparse matrix for noise reduction: the noise in the original data and the sample label data have been classified. The sparse expression function of the sparse matrix is ​​used to perform a noise reduction process on the noise data in the image; The dictionary learning sparse expression model is as follows: In the formula, X represents the sparse expression result image; I represents the image output after Bayesian classification; s represents the number of patches the image is divided into; Represents the transposed matrix of the image path matrix; E s represents the extracted image path matrix; Z represents the noisy image data; D represents the sparse expression matrix; α s Represents the sparse expression signal data of the patch image; (1.5) Perform image processing on the denoised data; such as image recognition and segmentation based on the denoised data; The image denoising method based on naive Bayes posterior probability and dictionary learning sparse matrix, through the powerful classification function of the naive Bayes model, finely classifies the noise data and sample label data in the image, and then provides the classification results to the dictionary learning sparse matrix for sparse expression, and achieves the effect of noise reduction and denoising through mathematical methods, thereby improving the signal-to-noise ratio of medical image data, as described in the following formula: C=Blb(1+S / N) Among them, C represents the maximum information rate that can be transmitted by the channel, that is, the transmission channel, referred to as channel capacity; B represents the bandwidth of the channel, that is, the transmission channel; S / N represents the signal-to-noise ratio of the received signal, and the core is to obtain higher quality image data by improving S / N; lb represents the logarithm with base 2.

Images