Research On Marine Microalgae Recognition Algorithm For Few-Shot Scenarios

In the marine ecological environment,marine microalgae are photosynthetic autotrophic organisms that occupy an important position.They are capable of performing photosynthesis and absorbing carbon dioxide.With the increasing severity of eutrophication in bodies of water,certain types of algae in the water reproduce rapidly and gradually form harmful algal blooms,destroying the water environment.Therefore,in order to maintain the ecological health of water and make full use of algae resources,it is very important to identify algae.In the marine microalgae image recognition task,manual microscopy recognition is labor-intensive,and traditional image recognition algorithms have problems such as long recognition time and low accuracy.Currently,deep learning algorithms are widely used for algae identification,although they have higher accuracy,they are complex and have a large number of parameters,and the recognition type is few,the goal is single.The extremely small volume and similar shape of marine microalgae make sample acquisition difficult.Based on the current research status of marine microalgae recognition both domestically and internationally,analyzing the independently constructed marine microalgae microscopic image dataset,it was found that it meets the characteristics of a few-shot dataset,i.e.,the sample quantity is small and the features are not apparent.Therefore,mainstream algorithms applicable to few-shot scenarios were chosen,namely,principal component analysis algorithm and few-shot learning algorithm,for recognizing and classifying marine microalgae.The research is mainly divided into two scenarios: recognition and classification of few-shot of similar microalgae images and recognition and classification of few-shot of whole multi-class microalgae images.The research content is as follows:(1)For the recognition problem of similar microalgae images in a few-shot scenario,a principal component feature-based similar marine microalgae few-shot recognition algorithm was proposed.Some of the algae in marine microalgae have similar morphologies,are difficult to distinguish,and have indistinct characteristics.Based on the traditional principal component analysis algorithm,this algorithm combines it with a logistic regression model at the feature extraction stage by calculating the principal component load of similar microalgae images and fusing them into the logistic regression model for classification.Experimental results show that under the same experimental conditions,the image recognition rate in this study increases by1.86% compared with the traditional support vector machine algorithm,while the accuracy reaches 92.86%,which effectively solving the recognition and classification problems of similar microalgae images in few shots.(2)For the recognition problem of whole multi-class microalgae images in a few-shot scenario,a marine microalgae few-shot recognition algorithm based on gridded multi-scale local feature fusion was proposed.Firstly,the input marine microalgae microscopic images are processed with gridded multi-scale processing and sent to the class traversal module for feature extraction to obtain local features.Secondly,a local feature fusion module based on self-attention mechanism is added to obtain locally enhanced features containing global information to improve the model’s generalization ability.Finally,the recognition and classification of marine microalgae is achieved by calculating the Euclidean distance between the feature vectors of the query set and the class prototypes in the support set as the metric.Experimental results show that under the experimental settings of 5-way 1-shot and 5-way 5-shot,the recognition accuracy of the proposed algorithm on the marine microalgae few-shot dataset is improved by 6.08% and 5.5%,respectively,which compared to the optimal recognition accuracy in the other eight few-shot learning mainstream recognition networks.It effectively solves the recognition and classification problems of whole multi-class microalgae images in few shots.This marine microalgae recognition algorithms based on few-shot learning provides a new path for algal recognition research.