Research On Feature Extraction And Recognition Of Deep-Sea Biological Image Based On Fractal Dimension And Convolutional Neural Network

The study of species,density and distribution of deep-sea organisms is of great significance for the exploration of gas hydrate,the comparison of biodiversity in different regions,the maintenance of deep-sea ecological balance,and the subsequent evaluation of resource exploitation and environmental impact.However,due to the large number,high density and low characteristics of deep-sea organisms,it is time consuming and low accuracy to recognize biological species and count biological numbers by using artificial recognition methods.Aiming at the problem of how to efficiently and accurately detect deep-sea organisms,this paper carries out feature extraction and recognition of deep-sea organisms images based on fractal dimension and convolutional neural network.The main research work is as follows:(1)To solve the problem that there is no publicly accepted deep-sea biological image dataset and related deep-sea biological image recognition research at home and abroad and,biological image data sets of deep-sea cold-spring and deep-sea hydrothermal are constructed from scratch through image labeling and file format conversion.To solve the problems of unbalanced species,insufficient quantity and low quality of of deep-sea biological image samples,as well as over-fitting and poor generalization ability in model training,the geometric transformation method and image augmentation method are used to expand the existing training samples.The results show that by making new datasets and expanding existing training samples,data augmentation is realized,which can effectively expand the number of images and improve the quality of images.(2)Since the fractal dimension can describe the texture roughness of the image surface,an image classification algorithm based on box-counting dimension feature extractor and OVO-SVM classifier is proposed to complete the task of feature extraction and classification of a single deep-sea biological image.By designing the feature extraction algorithm based on the box-counting dimension,the approximate box-counting dimension of four kinds of single deep-sea biological images are calculated and input into OVO-SVM classifier as feature vectors for training.Finally,the specific categories of test samples are determined by voting mechanism.The results show that the proposed method performs well in the task of feature extraction and classification of a single deep-sea biological image.(3)According to the characteristics of deep-sea dense biological images,an image recognition method based on convolutional neural network is proposed for deep-sea cold-springs and deep-sea hydrothermal intensive biological image recognition.By analyzing the principles and structures of three feature extraction networks and two types of object detection algorithms,six deep learning models have been constructed and improved.The optimal model is selected by weighing the average detection speed,average confidence score and other aspects,and then relevant application researches are carried out according to this optimal model.The results show that the proposed method has a better recognition effect and a higher classification and location accuracy in the deep-sea dense biological image recognition,and the related application researches are feasible and have practical value.