CN-121392352-B - Fuzzy monotone correlation image recognition and machine learning method

Abstract

The invention discloses a novel fuzzy monotone correlation image recognition and machine learning method, which belongs to the technical field of artificial intelligence of pattern recognition and machine learning, and defines the method as FMMCA, wherein local fuzzy monotone correlations are evaluated by comparing row vectors and column vectors of an image matrix one by one, and then the local correlations are weighted and accumulated to finally obtain global fuzzy monotone correlations among images. The invention directly uses fuzzy monotone correlation analysis to replace classical correlation analysis to carry out multi-view research, so that the problems of classical correlation analysis can not exist, the characteristics of the fuzzy monotone method do not need to carry out static measurement by distance, and can offset some noise influence by interval change dynamic measurement, thereby improving the performance, forming a novel fuzzy monotone machine learning method, and the invention does not need to optimize a focusing loss function, has few parameters, good robustness and little calculation power.

Inventors

• LIANG JIN

Assignees

• 华南师范大学

Dates

Publication Date

20260508

Application Date

20250910

Claims (4)

1. 1. A fuzzy monotonic correlation image recognition method, the method comprising the steps of: s1, giving two data sets And Where N represents the sample size of the image, Is the i-th sample image of X, Is the jth sample image of Y, i is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to N, and X i and Y j are respectively marked as a matrix A m×n and a matrix B m×n ; S2, extracting correlation characteristics between X i and Y j , respectively extracting corresponding column vectors and row vectors from a matrix A m×n and a matrix B m×n to perform correlation calculation, and integrating local correlation results by a weighted accumulation mode, wherein the method specifically comprises the following steps: S2.1, calculating the local correlation in the row dimension by traversing all row vectors of the matrix A m×n and the matrix B m×n , wherein the mathematical expression is as follows: wherein A T represents the transposed matrix of A, Represents column i of a T , B T represents the transposed matrix of B, Column i of B T , m the height of the sample image, k 1 the number of intervals, p 1 the number of elements within an interval, The relationship between k 1 and p 1 is shown below: where n represents the width of the sample image, when p 1 = 1, Is divided into n intervals, and each interval has only one element; The parameter t 1 is the optimal upper bound of k 1 , the parameter t 1 is gradually adjusted by adopting an iterative optimization method, and an optimal t 1 value is finally found by evaluating the accuracy of image classification after each adjustment; S2.2, calculating the local correlation in the other dimension by traversing all column vectors of the matrix A m×n and the matrix B m×n , wherein the mathematical expression is as follows: Where the parameter t 2 is the optimum upper bound for k 2 , A j represents the j-th column of A, B j represents the j-th column of B, k 2 represents the number of intervals, p 2 represents the number of elements within an interval, The relationship between k 2 and p 2 is as follows: S2.3、 And The overall correlation between is obtained by combining the two dimensions of rows and columns, expressed as follows: Wherein, the The higher the value of (2), the description And The stronger the fuzzy monotonic relation strength between the two sample images, the stronger the correlation between the two sample images; s2.4, during machine learning training, training an optimal combination of t 1 and t 2 according to the classification accuracy, namely t, setting t 1 and t 2 to be equal to a certain t uniformly, training, and finding out interval division t with highest classification accuracy as a machine training feature; s3, constructing an image frame to realize image classification tasks, and completing image classification by comparing correlation strengths among images.
2. 2. The method for identifying a fuzzy monotonic correlation image according to claim 1, wherein S3 specifically comprises the following contents: assume that the image class of sample image Y r is known, and And finding Y r with the largest correlation with X i through iterative updating so as to classify X i and Y r into the same image category, wherein the mathematical formula for completing the image classification task is expressed as follows: Where r represents the number of iterations.
3. 3. A computer device comprising a processor and a memory having stored therein at least one instruction, at least one program, code set, or instruction set that is loaded and executed by the processor to implement the blurred monotonic correlation image recognition method as recited in any one of claims 1-2.
4. 4. A computer readable storage medium having stored therein at least one instruction, at least one program, code set, or instruction set loaded and executed by a processor to implement the blurred monotonic correlation image recognition method as recited in any one of claims 1-2.

Description

Fuzzy monotone correlation image recognition and machine learning method Technical Field The invention relates to the technical field of artificial intelligence of pattern recognition and machine learning, in particular to a fuzzy monotone correlation image recognition and machine learning method. Background Typical correlation analysis (Canonical Correlation Analysis, CCA) has received widespread attention since the proposal in 1935 by h. The goal of CCA is to extract the best predictor from each set by maximizing the correlation coefficient between the two sets. To achieve this goal, hotelling further proposes a symmetrical solution, defining the sequence of variable pairs as typical variables, and the correlation between them is typically the correlation. By projecting a complex set of high-dimensional variables into the desired low-dimensional common hidden subspace, CCA is able to extract the maximum correlation between two sets while simplifying the statistical analysis of the two set variables. In recent years, CCA has become a hotspot in research areas and has found widespread use in a number of discipline areas, including computer vision, biomedical, natural language processing, genetics, knowledge and data engineering, and the like. However, while CCA plays a critical role in multi-view correlation analysis and plays a key role in research fields such as computer vision and pattern recognition, there are some limitations in its application. In particular, CCA relies on Pearson correlation coefficients as a measure of correlation, which may not adequately capture the nonlinear relationship between data when analyzing the correlation between data. Furthermore, when the data set contains noise, the performance of the CCA may be significantly affected because the noise data may distort or mask the true correlation pattern. Thus, CCA may be limited in its effectiveness in processing non-linear data and noisy data, requiring further research and improvement to overcome these limitations. Over decades of development, a number of CCA-related algorithms have been proposed. As an extension of CCA, a core canonical correlation analysis (Kernel Canonical Correlation Analysis, KCCA) can find a nonlinear projection of maximum correlation, but is limited to a regenerated core hilbert space with corresponding cores. By nuclear trick KCCA the non-linear data is mapped to a higher (even infinite) dimension of the hilbert feature space where the data exhibits strong linearity. In 2014, randomized nonlinear canonical correlation analysis (Randomized Nonlinear Canonical Correlation Analysis, RCCA) was proposed. The study suggests that the construction of the randomization approach helps reveal the characteristics of the nonlinear pattern in the data. For basic tasks such as regression or classification, compared with an accurate kernel method, the nonlinear random characteristic is used, so that almost no loss in performance is caused, and the computational complexity is effectively reduced. In recent years, deep neural networks (Deep Neural Networks) have become a hotspot for research. Such networks are of great interest due to their excellent performance in a variety of tasks. One significant advantage of deep neural networks is that they can learn complex nonlinear representations of data, which makes them potentially enormous in many fields of image recognition, speech processing, and natural language processing. Furthermore, deep neural networks do not rely on traditional non-parametric models, meaning that they can automatically adjust parameters to accommodate different data sets and task requirements. In view of these advantages of deep neural networks, many researchers have begun exploring methods that combine deep learning with CCA. This combination takes advantage of the automatic feature extraction capability of deep neural networks and the correlation analysis capability of CCA, with the aim of learning a deeper, richer representation of the data. For example, andrew et al propose depth canonical correlation analysis (Deep Canonical Correlation Analysis, DCCA) that mines more complex nonlinear correlations between two views through a deep neural network. Wang et al propose a typical correlation analysis based on a depth self-encoder (Deep Canonical Correlation Autoencoder, DCCAE), which model consists of two self-encoders and minimizes the reconstruction error of the self-encoders under the CCA framework. Benton et al propose a deep generalized canonical correlation analysis (Deep Generalized Canonical Correlation Analysis, DGCCA) and apply it to three tasks of speech transcription, pronunciation measurement and information recommendation. Chandar et al propose a correlated neural network (CorrNet) further coupled with cross-view reconstruction errors, thus providing an accurate reconstruction for one view given another view. Yang et al propose a convolutional neural network-based canonical correl