CN-119295799-B - Colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm

Abstract

The invention belongs to the technical field of image processing, and discloses a colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm. The method comprises the steps of collecting a dataset, screening and clipping collected image data, extracting characteristics of pictures by using PERSISTENT HOMOLOGY, introducing an ECA attention mechanism into a MBConv module in a EFFICIENTNET-B0 classification model, and classifying images secondarily based on an SCoT self-attention mechanism. The invention realizes higher recognition accuracy with less parameter quantity.

Inventors

• WANG ZUMIN
• YANG KE
• HUANG JUNMING
• ZHANG YUHAO
• GAO JUN
• YAN HUAN

Assignees

• 大连大学

Dates

Publication Date

20260512

Application Date

20240911

Claims (3)

1. 1. A colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm, characterized by the steps of: s1, collecting a data set and screening and cutting the collected image data; s2, extracting features of the picture by using PERSISTENT HOMOLOGY; s3, introducing an ECA attention mechanism into a MBConv module in a EFFICIENTNET-B0 classification model; S4, performing secondary classification on the images based on an SCoT self-attention mechanism; The step S3 specifically comprises the following steps: S3.1, inputting data to an improved EMBConv structure, namely inputting the image data processed by common 3x3 convolution kernel into a EMBConv structure module, wherein the image data is subjected to PH topological feature extraction and data enhancement; S3.2, introducing a1 multiplied by 1 convolution kernel; s3.3 leading in 3X 3Depthwise Conv convolution, wherein each input channel is processed by using an independent 3X 3 convolution kernel, the number of channels is kept unchanged, then batch normalization is applied to normalize the data of each channel to have a distribution with a mean value of 0 and a standard deviation of 1, and then nonlinearity is led in through Swish activation function; S3.4, entering an ECA module, namely identifying key features and important information in the data through an ECA attention mechanism, giving higher weight to the information, and neglecting relatively unimportant information; s3.5, leading into a common convolution of 1x1, comprising a BN module; s3.6, entering Droupout layers; S3.7, importing output data into a remaining six-layer MBConv module of the neural network, and inputting the output data into an SCoT self-attention mechanism; the step S4 specifically includes: S4.1, inputting data, namely importing the data subjected to the pre-processing into an SCoT self-attention mechanism; S4.2, dividing the feature map S into N parts, representing the N parts by S 0 ,S 1 ,…,S n-1 , wherein the number of channels of each divided part is C1=C/N, and simultaneously, for each divided channel feature map, extracting the space information of different scale feature maps by multi-scale convolution kernel Group convolution, and adaptively selecting the Group size according to the size of a convolution kernel; S4.3, a spatial branch part processing, in which the self-attention block operates on the input tensor X to highlight or suppress the characteristics, the dynamic range of attention is increased by Softmax normalization on the bottleneck tensor, and tone mapping is carried out by using a Sigmoid function; S4.4, for the combined context information branch learning part, for the feature modules of two-dimensional different channels formed by the segmentation in the previous step, representing each key by using k x k group convolution for all adjacent keys in the k x k grid in space and combining context content; S4.5 combining two learning methods, taking the concatenation of the static context key K 1 and the query Q as a condition, implementing the attention matrix by two consecutive 1X 1 convolutions, including those with ReLU activation functions And no activation function ; S4.6, outputting the processed data to a1 multiplied by 1 convolution layer and a pooling layer for processing, classifying the features by adopting a linear classifier Softmax, and outputting a recognition result.
2. 2. The colony classification method based on PERSISTENT HOMOLOGY and modified EFFICIENTNET algorithm according to claim 1, wherein said step S1 is specifically: s1.1, collecting a data set, namely, using the data set as a self-collecting data set of candida albicans and staphylococcus epidermidis; S1.2 screening, namely screening out the colony parts which are incomplete in adhesion and growth and are incomplete in shooting, and reserving complete single colony data of the colony parts; S1.3, cutting, namely cutting single bacterial colonies with proper sizes for the bacterial colony pictures reserved after screening by comprehensively considering, and enlarging the data set by a data enhancement method for the data set subjected to screening and cutting treatment.
3. 3. The colony classification method based on PERSISTENT HOMOLOGY and modified EFFICIENTNET algorithm according to claim 1, wherein said step S2 is specifically: S2.1, constructing a point cloud, namely generating a set P containing data points from a provided picture, wherein the points are called the point cloud, and the points are points in Euclidean space or other measurement space; s2.2 determining parameters, selecting one parameter Representing the size or radius of the build shape; S2.3 construction of the complex shape, use to Is of radius The dashed sphere surrounds each point in P and connects between this point and all other points in its circle; s2.4, analyzing the topological structure, namely acquiring topological information about the data set by analyzing the constructed complex topological structure.

Description

Colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm Technical Field The invention belongs to the technical field of image processing, and particularly relates to a colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm. Background In recent years, with the development of computer technology, image processing technology and artificial intelligence theory and technology, computer vision technology has been rapidly developed and widely used, so that the computer vision technology plays an important role in the scientific research field and the medical direction army valve. The computer vision technology is a technology with extremely high comprehensiveness, relates to the subject fields of artificial intelligence, optics, mechanics, computer graphics, neurobiology and the like, has been widely applied to the aspects of business, military, weather, natural disaster prediction and the like, and brings the change of over the sky and over the earth to the life of people. Technologies such as face recognition and unmanned driving are not supported by computer vision. Meanwhile, computer vision technology has been developed from single dimension to multi-dimension in continuous updating iteration, and the related fields are also wider and wider. The application of the computer vision technology in the medical field is mainly embodied in aspects of influence identification and analysis, pathological image analysis, operation coaching and navigation, health detection and remote diagnosis, personalized medical treatment and the like. Most of the diseases in today's society are caused by bacterial or fungal infections, such as respiratory tract infections, urinary tract infections, skin infections, etc. Although some pathogenic bacteria are different, the infection symptoms are almost consistent, such as dermatitis or acne caused by candida albicans and staphylococcus epidermidis skin infection, and hyperpyrexia caused by blood infection, which is difficult for common people to distinguish. For this, rapid antigen, PCR detection and colony culture are often used to identify the species. However, the rapid antigen method and PCR detection have a certain error in the results although the speed is high, and the colony culture result has high accuracy but long time. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm, which is divided into two parts of feature extraction and classification treatment, wherein in the data preprocessing stage, PH (Persistent Homology) is used for strengthening key feature areas of images and amplifying datasets. The MBConv (Mobile Inverted Bottleneck Convolution) module in EFFICIENTNET is redesigned according to the design idea of the convolution network, and the ECA convolution attention is introduced after the first convolution. And introducing a new self-attention mechanism after the last layer of MBConv modules, namely, performing multidimensional feature weight distribution learning by the SCoT self-attention mechanism module, and finally classifying the features by adopting a linear classifier Softmax and outputting a recognition result. The aim of the invention is achieved by the following technical scheme that the colony classification method based on PERSISTENT HOMOLOGY and improved EFFICIENTNET algorithm comprises the following steps: s1, collecting a data set and screening and cutting the collected image data; s2, extracting features of the picture by using PERSISTENT HOMOLOGY; s3, introducing an ECA attention mechanism into a MBConv module in a EFFICIENTNET-B0 classification model; s4, performing secondary classification on the images based on the SCoT self-attention mechanism. Further, the S1 specifically is: S1.1 data set acquisition the use data set was a self-acquired data set of Candida albicans and Staphylococcus epidermidis. S1.2 screening, namely screening out the colony parts which are incomplete in adhesion and growth and are incomplete in shooting, and reserving complete single colony data. S1.3, cutting, namely cutting single bacterial colonies with proper sizes for the bacterial colony pictures reserved after screening by comprehensively considering, and enlarging the data set by a data enhancement method for the data set subjected to screening and cutting treatment. Further, the S2 specifically is: s2.1, constructing a point cloud, namely generating a set P containing data points from a provided picture, wherein the points are called the point cloud and can be points in Euclidean space or points in other measurement spaces; S2.2 determining parameters by selecting a parameter epsilon representing the size or radius of the build shape. Epsilon determines the maximum distance between two points which can form connection, and topological feature extrac