High Density Shrimp Seedling Counting Method Based On Deep Learning

Since the 21 st century,China has entered a phase of rapid development,in which the rapid development of e-commerce and logistics and transportation modes have made it possible for people in all regions of China to purchase fresh seafood,which has led to an increasing domestic demand for seafood such as fish,shellfish and shrimp.In this situation of oversupply,the use of aquaculture as well as improving the efficiency of aquaculture has become particularly important.Accurate and efficient farming methods help to produce more and better shrimp products.Accurate estimation of the number of shrimp fry in farming waters is of relevance for shrimp fry farming,shrimp fry trading and shrimp fry behavior analysis.In the case of shrimp farming,the most critical step is the screening and counting of shrimp fry.Traditional fry screening and counting is based on direct observation,i.e.,relying on expert visual observation of fry.The traditional screening and counting method is simple,but it is costly in terms of time and labor and is not very accurate.Therefore,the use of direct observation does not guarantee that the task of counting the number of shrimp fry can be done effectively for a long time.Computer vision technology has been rapidly developed in recent years,including applications to face recognition and crowd counting,but the application to the shrimp fry counting problem still faces more challenges.The use of accurate and efficient counting methods for shrimp fry has become one of the urgent problems in the field of smart fisheries.In order to solve the above problems,this paper proposes two deep learning-based shrimp fry density estimation methods based on an in-depth study of existing shrimp fry counting methods.The specific work focuses on the following points.1.Considering the insufficient existing shrimp fry dataset situation,we collected and labeled a dataset(Dlou＿Shrimp)for shrimp fry counting,which collected and organized 250 images and did accurate labeling of shrimp fry in about 80,000 images,which provided good data assurance for the subsequent experiments.2.To address the problems of scale variation and feature loss in shrimp fry density estimation,a shrimp fry density estimation algorithm based on multi-scale dual-channel fusion is proposed,which consists of Res Net101 as the feature extraction module of the whole network;then after the multi-scale dual-channel fusion module with SKNet,by letting the network self-assign the scale of the convolution kernel,the model is allowed to learn regions of different densities with targeted learning to cope with background noise and uneven distribution of shrimp fry;finally,attention map loss is also introduced to compensate for the Euclidean loss with more explicit supervisory information.The experimental results demonstrate that the multi-scale dual-channel fusion algorithm designed in this paper is able to learn the multi-scale features of shrimp fry well,thus reducing the counting error and improving the performance of the model.3.For the uneven distribution of shrimp fry itself and the problem of occlusion and adhesion,we propose an improved Unet-based shrimp fry density estimation method.Considering better learning of contextual and location information,an improved attention mechanism module is added to the feature map fusion phase of the original Unet network to improve the focus on the dense shrimp fry areas.For the problem of uneven distribution of the binary matrix existing between geometric adaptive Gaussian kernel and fixed Gaussian kernel,the loss function of DMCount is used instead of the original loss function in this paper.Experiments show that the method proposed in this paper can solve the problem of occlusion and adhesion in shrimp fry counting and provide a new idea for shrimp fry density estimation.