| At present,the methods of counting shrimp fry on the market mainly adopt manual methods such as photoelectric counter,standard dry container method and weight method.Due to the small size,high density,and easy accumulation of the shrimp fry,these counting methods are prone to inactivate the shrimp fry during the emergence stage and cause economic losses,and it is complicated.However,as the total production of shrimp fry in China has increased year by year,it has given birth to the urgent development of the industrialization of fry counting.Therefore,there is an urgent need for a technology that is simple to operate and can automatically count shrimp fry.Recently,deep learning has many applications,the convolution neural network is likely to fill in aquaculture shrimp larvae automatically enumerate blank.This paper will use convolutional neural network to study a technology of automatic shrimp fry counting.The main tasks and innovations completed are as follows:1.After reading a large number of domestic and foreign literatures,the method of labeling shrimp larvae using pixel markers was first proposed.The center of the shrimp fry head is marked with a pixel point as the correct result(ground-truth)of the sample object,and then a high-quality density map is generated to count the shrimp fry through a two-dimensional Gaussian kernel function.2.By improving the basic network of VGG16,an estimation algorithm based on the density distribution of shrimps is designed.The specific method is as follows:(1)Remove the fully connected layer of the basic network and use the convolutional layer instead.(2)Reduce the convolutional layer in the front part of the network,while effectively fusing the feature maps generated by different convolution kernels.(3)Replace the fifth pooling layer of the basic network with a 1 × 1 convolution kernel to process the output feature map.(4)Use dilated convolution to deal with the receptive field mismatch problem caused by the removal of stride in the fourth largest pooling layer.(5)Input the shrimp data training set and the real density map to the network model to learn the mapping relationship from the image feature to the density distribution map,and integrate the distribution density map to calculate the final number of shrimp fry.3.Under the premise of using the same equipment to shoot the shrimp fry at the same time period,a total of 312 original images of shrimp fry with different densities and lengths of about 6to 15 mm were collected.851 image samples were obtained after rotation,flip,translation and other enhancement processing,and it took 137 hours to use different manual marking methods,and finally a total of 739 519 shrimp larvae were labeled as experimental samples.4.Use the same data set to explore experiments with different labeling methods on algorithms such as YOLO and SSD.At the same time,compared with the classic counting network,the results show that compared with MCNN,CSRnet,CAN,the average absolute error of this method can be reduced by 7.6,4.8,3.2,respectively,and the root mean square error can be reduced by 8.4,6.2,3.4,respectively.The accuracy rate is as high as 97.59%.It has been verified that the algorithm in this paper can accurately estimate the number of shrimp fry of different densities in a uniform backlight environment,and has higher robustness and stability.It solves the problem of counting shrimp fry with a certain density and overlap degree and reduces the inconvenience caused by manual manual counting of seedlings,which meets the counting requirements of the shrimp fry industry. |
