Algorithm Design And Embedded System Implementation For Marine Zooplankton Image Recognition

Marine zooplankton is widely distributed in the ocean,and closely related to human life.In recent years,the development of in situ plankton imaging technology has greatly improved the efficiency of in situ image data acquisition.Meanwhile,the application of computer and statistical analysis technology to automatic recognition of the ocean images has also attracted wide attention and shown good results in many close-set plankton classification datasets.However,the recognition of zooplankton in the natural seawater is still facing many challenges: the composition of the seawater is extremely complex and difficult to be exhausted,and the recognition inevitably encounters open-set problems;the distribution of zooplankton is naturally uneven and is changing dynamically;a large amount of non-planktonic particles will also interfere with zooplankton recognition,etc.These factors may lead to a significant decrease of the recognition performance in practice.Based on that,this work aims to achieve zooplankton recognition in natural seawater,and conducts research on in situ image preprocessing,zooplankton image recognition,and the implementation of recognition algorithms on an embedded platform,respectively.In the in situ image preprocessing,a preprocessing pipeline is proposed to extract the targets that meet the observation requirements.An algorithm based on the sliding window is proposed for the underwater object localization,which achieves good results in locating low-brightness targets under high turbidity seawater.For the focusing evaluation of zooplankton images,this work proposes an algorithm based on edge gradient feature,and an algorithm based on convolutional neural network,respectively.They complement each other in computational efficiency and evaluation performance.To solve the problems of data imbalance,dataset drift,and open-set problems which may face in natural seawater,this work proposes to use image retrieval for zooplankton image recognition.The algorithm achieves excellent performance under the combined effect of contrastive learning and good image quality of darkfield color images.Besides,it is also more flexible and convenient than common classification methods,and can meet the needs of several in situ observation applications.Finally,on the basis of the algorithms proposed,this work implements the algorithm on an embedded platform.The implementation of in situ image processing program and the acceleration of the neural network deployment are completed,which effectively avoids the pressure of data transmission,and meet the demand of in situ real-time observation.