摘要： 红螯螯虾人工繁育中需要对离体虾卵进行准确计数。由于虾卵个体小、密度大，人工计数效率低且可能损伤虾卵，提出了一种基于残差模块的计算机视觉虾卵计数模型。该计数网络模型在前端编码器网络中集成残差模块，对虾卵位置信息进行特征提取和下采样，在后端解码器网络中使用转置卷积，使模型可以自动学习上采样参数。与VGG-16、CSRNet等模型相比，基于残差模块的虾卵计数模型精度达98.88%，具有最低的平均绝对误差、均方误差以及最好的预测密度图质量，可有效解决密集虾卵计数的困难，为红螯螯虾虾卵计数提供了新方法。利用该模型的计算机视觉虾卵自动计数装置，结构简单、计数精准，可以应用于真实场景。

关键词: 红螯螯虾, 虾卵计数, 残差模块, 密度图, 深度学习, 计算机视觉[]

Abstract:  It is necessary to count the shrimp eggs in vitro accurately in the artificial breeding of red claw crayfish. Due to the small size and high density of shrimp eggs, the artificial counting of shrimp eggs has low efficiency and may damage the eggs. We propose an end-to-end shrimp eggs counting model that is based on the residual building unit. This network model integrates a custom residual building unit in the front-end of the encoder network and performs feature extraction and down-sampling of the shrimp eggs’ position information. The back-end of the decoder network uses the transposed convolution instead of bilinear interpolation, so that the model can learn upsampling parameters automatically. Experiments on the shrimp eggs dataset, compared with VGG-16, CSRNet, MCNN, SANet, the proposed counting method based on the residual building unit can reach the counting accuracy of 98.88%, and has the lowest mean absolute error, the lowest mean square error, and the best quality of the predicted density map. These test results show that the proposed counting method of red claw crayfish’s eggs based on a residual building unit can effectively solve the difficult counting problems caused by the small size, high density and easy adhesion of the shrimp eggs, and provide a new method for the counting of red claw crayfish’s eggs. The computer vision automatic counting device for the shrimp eggs was based on the proposed counting model, which had the characteristics of simple structure and accurate counting, and could be applied to real scenes.

Key words: red claw crayfish, shrimp eggs counting, residual building unit, density map, deep learning, computer vision