Multi-objective Recognition And Color Restoration Of Coral Community Based On Deep Learning

Human activities and global warming have led to the degradation of coral communities and faced a survival crisis.In order to prevent the irreversible damage and degradation of coral communities,the early monitoring of coral communities is very important.At present,the monitoring and evaluation methods of coral communities in China are relatively backward.Under this background,the research on multi-objective recognition and color restoration methods of coral communities based on deep learning is of great significance to improve the monitoring and evaluation methods of coral communities,prevent the degradation of coral communities It is of great significance to implement the national strategy of "maritime power" and the construction of marine ecological civilization.Coral community monitoring is a method to systematically evaluate the biodiversity and health status of coral communities.Through the identification and evaluation of reef organisms in coral communities,we can systematically judge the state of biodiversity of coral communities,so as to judge whether the living conditions of coral communities as ecosystems have been destroyed;Through the color restoration of coral image,we can judge whether there is albinism or disease through the color of coral to evaluate its health status.Therefore,the multi-objective recognition of reef organisms in coral community and the color restoration of coral image are the two most basic and important problems in coral community monitoring.The research on multi-objective recognition and color restoration methods of coral community based on deep learning is still in its infancy,so there are still some problems: on the one hand,for the multi-objective recognition of reef organisms in coral community,the complex imaging environment of real seabed scene leads to the fuzzy characteristics of coral colony reef organisms,and most reef organisms are small targets and clustered,It further increases the difficulty of multi-target recognition of underwater coral reef ecosystem.On the other hand,for the problem of coral image color restoration of coral community,coral community is mostly located in the seabed reef bed.The complex environment leads to the changeable degradation of coral image,and the effect of coral color restoration is not ideal.At present,the collection of coral community image is still manual,and the paired image data is scarce,which increases the difficulty of coral image color restoration.For the above two problems,this paper proposes the following solutions:1)Firstly,aiming at the problem of reef habitat feature blur caused by the degradation characteristics of real offshore images,yolov5 algorithm is introduced,and the jump connection operation is designed to directly transfer the shallow feature information to the deep network,so as to effectively reduce the impact of deep network feature blur on recognition accuracy,Thus,the recognition accuracy of reef organisms is effectively improved without adding additional calculation.Secondly,in order to reduce the influence of invalid features in the jump link on network training,we introduce the Convolutional Block Attention Module(CBAM)and embedded in the jump connection process.The module will input the feature graph to make attention inference,and then adaptively advanced the input features.The proposed improved algorithm can further improve the recognition accuracy of reef organisms in coral community without additional computational cost.2)In view of the poor color restoration effect of underwater coral images and the scarcity of paired coral community images,firstly,a coral image color restoration model based on unpaired coral image data set is proposed,which solves the problems of few training images and poor training effect due to the difficulty of acquiring paired data.At the same time,aiming at the problems of stopping updating parameters and easy over fitting in the training of the network,the Silu activation function and do conv convolution module are introduced to improve the effect of feature extraction during training.Secondly,an unpaired coral image data set composed of Hong Kong sea coral image and network coral image is established for training,which increases the image data of training,Provide reliable support for follow-up research.Through the experiment of effectiveness and advanced,it is proved that the improved algorithm can effectively restore the true color of coral image.