Abstract: In the process of injecting the virus into the egg embryo to culture the vaccine, the impact of the needle caused cracks around the injection hole and  the culture failed. In order to solve the problem of low efficiency and high misjudgment rate in the detection of egg embryo cracks by the naked eye in the darkroom, a method for detecting cracks in egg embryos with holes combined with a multispectral channel attention mechanism (MSCA) and parallel stacking topology modules (PST) was proposed. Firstly, a black box for crack detection of perforated egg embryos was built, and the images of linear cracks, mesh cracks, starshaped cracks and intact egg embryos with holes after virus injection was collected and a data set was established, then ResNet-50 was used as the backbone model. The first residual module of the next 4 layers was replaced by the PST module to increase the image expression ability of the model at the beginning, finally, the MSCA mechanism was introduced after each PST module and the residual module, and the MSCA mechanism compressed the data through twodimensional discrete cosine transform (2D DCT),  and the neural architecture search selection (NAS) method obtained the optimal frequency components, which could redistribute the weights to increase the proportion of crack feature weights. Thus the microporous egg embryos could be identified quickly and accurately by the model. The results show that the improved network model has a detection time of 0.42 s/piece for cracks in egg embryos with holes, a detection accuracy of 96.43%, and a higher detection efficiency than manual operations. Compared with the original ResNet-50, the detection accuracy has been improved by 3.66%, which is superior to other classical convolutional network models. It is proved that the improved model can be used for the automatic detection of cracks in egg embryos with holes before vaccine cultivation.

Key words: crack detection, egg embryo with hole, improved ResNet, multispectral attention mechanism, parallel stacking topology

摘要： 将病毒注入蛋胚培养疫苗时，针头冲击使注射孔周围产生裂纹导致培养失败。为解决目前人工手持照蛋器在暗室通过肉眼检测蛋胚裂纹效率低、误判率高这一问题，提出一种结合多光谱通道注意(MSCA)机制与并行堆叠拓扑(PST)模块的带孔蛋胚裂纹检测方法。首先搭建带孔蛋胚裂纹检测黑箱，采集病毒注射后带孔蛋胚的线形、网状、星形裂纹和带孔完好蛋胚图像并建立数据集；接着以ResNet-50为骨干模型，将其后4层的第一个残差模块替换为PST模块，以增加模型初期图像表达能力；最后在每个PST模块与残差模块后引入MSCA机制，MSCA机制通过二维离散余弦变换(2DDCT)压缩数据得到各通道频率分量，采用神经结构搜索(NAS)方式得到最佳频率分量，即对权重重新分配，增加裂纹特征权重比例，以确保模型快速、精准识别带孔蛋胚裂纹。结果表明，改进的网络模型对带孔蛋胚裂纹检测时间为0.42 s/枚，检测精度为96.43%，检测效率高于人工作业。与原始ResNet-50相比，检测精度提高3.66%，优于其他经典卷积网络模型，改进模型可用于疫苗培育前带孔蛋胚裂纹自动检测。

关键词: 裂纹检测, 带孔蛋胚, 改进残差网络, 多光谱通道注意机制, 并行堆叠拓扑模块