Development Of Large-scale Marine Fish Image Database And Its Applications

There are more researchers engaged in marine research in China,but there are fewer image databases of large marine fish.With the wide application of deep learning in image classification,researchers have more and more requirements on marine fish images.But some scattered small data sets have the problems of less data samples,inaccurate classification and inconvenient query,which bring inconvenience to the relevant research work.In view of these situations,the purpose of this thesis is to development a large-scale database of marine fish images.The research contents are as follows.1.A fish image processing method based on cnn-4 neural network model was developed.We adopt the web crawler method to expand the data based on Fish Base and Wild Fish database.Image-aware hash algorithm is used to deduplication the repeated data collected.In order to improve the efficiency of data cleaning,the convolutional neural network model(CNN-4)was proposed to clean the mass data with noise.The experimental results show that the data cleaning through the cnn-4 network model is very effective.The quality of large-scale marine fish image database has been significantly optimized.2.An automatic fish image classification method based on VGG-16 neural network model was studied.VGG-16 model structure was designed according to the characteristics of fish image to improve the efficiency of image classification.This method can reduce lots of manpower and material resources cost in image processing.It also reduces the requirements for staff with the professional background in marine biology.Finally,the classification effect of vgg-16 model is verified by experiments with the image classification accuracy and recognition time as the comparison indexes.Experimental results show that this method can not only guarantee the accuracy of image recognition,but also shorten the time of image recognition.And the process of image classification have been accelerated.3.A large-scale image database system for marine fish was developed.The system is developed by combining the retrieval program with the HBase database.Users can get the corresponding image of marine fish through fuzzy search or accurate retrieval by inputting text into the system.Different from the existing marine fish database,this system combines the deep learning method to realize image recognition and image retrieval.The user can retrieve the category of fish and the image of similar fish by the way of entering the image of fish through the system.This system improves the efficiency of data searching.We finally established a large-scale database for real scene images of marine fish through the research of this thesis.The database contains 107,935 images of 3,766 species of marine fish currently.The quality of database is improved gradually through the methods of this thesis.The feature will meet the data needs of researchers,facilitate the research of automated Marine fish identification technology,and promote the research of deep learning in Marine biology.A large-scale image retrieval system for marine fish is developed based on the database.The system has realized the functions of accurate search and fuzzy search,and image recognition and image retrieval functions are realized by the way of deep learning.It provides a retrieval platform for the application of different requirements in scientific research,education,fishery production and so on.