Research On Shrimp Eggs Counting Method Based On Convolutional Neural Network

In recent years,the scale of shrimp farming has increased rapidly.And hence,there is a growing demand for intelligent technology related to artificial reproduction.Accurate counting of shrimp eggs in the hatchery is the basic requirement in the process of artificial hatch.Accurate quantitative information can provide a wealth of information feedback to the farm managers,such as the number of eggs held by female shrimp,hatching rate,survival rate,etc.These information can help the managers to achieve more detailed and effective stocking management.However,at present,shrimp eggs counting is mainly performed by manual operation in a time-consuming way,and the consistency between different operators is not guaranteed.On the other hand,deep learning is developing rapidly and has achieved excellent performance in image processing and image recognition.Compared with the traditional methods for target counting,the algorithms based on Convolutional Nueral Network(CNN)have great advantages in speed and precision.According to the development trend,studies and develops the CNN-based shrimp eggs counting methods,with considering some problems in the practice.The main research works are specifically as follows:Firstly,we describe the state of the art of the research on target counting in detail.The basic concepts and related principles of CNN and GAN are introduced.Also,we analyze and compare several classic CNN models.Secondly,in view of the lack of relevant datasets for shrimp eggs counting,we share a dataset of Australian red crayfish eggs,which includes 450 images with about 272,000 eggs annotated.In addition,we introduce a label-free method based on GAN to generate the synthetic dataset.The method eliminates the labor cost in the process of dataset labeling and improves the precision of dataset labeling.Thirdly,in order to maintain the low-level features about the eggs such as the position and the contour,we proposed a residual based shrimp eggs counting model(RECNet).The whole model adopts an encoder-decoder structure.In the encoder part,the architecture is based on the residual module,low-level information such as the position and contour of shrimp eggs can be retained.In addition,a VGG-16 based shrimp eggs counting model(SECNet)was designed.We directly carve the first ten convolution layers and three max-pooling layers from VGG-16 as the frontend to finetune.To get a larger receptive field,we use the convolutional layer and transposed convolution as the back end.Then,a series of comparison experiments were conducted on the real dataset and synthetic dataset.The experimental results show that the proposed method performs well in the shrimp eggs couting.Finally,based on the proposed methods,a simple automatic shrimp eggs counting system is designed and built.We test the counting system in the practice,and the result shows that this system using the algorithms proposed in this thesis are feasible and accurate.