An Innovative Algorithm to Process Imaging Data for Detection of Weak Chick Embryos in Vaccine Production

Abstract

Efficient and accurate detection of weak embryos often requires the detection and classification of both weak and live embryos over a specific period. In this work, we image, detect and classify weak and live embryos hatched between the 9th and 15th day. We introduce a new method called Weak Embryo Detection Network (WEDNet), which is a weak embryo detection method based on a multiscale-feature fusion convolution neural network. First, we broaden the expand convolution layer of the Fire module to implement a multiscale feature extraction using a lightweight network SqueezeNet structure, where convolution kernels of different sizes are stored. A residual connection method is introduced between adjacent modules to achieve feature fusion between layers. Furthermore, a residual multiscale Fire block (RMFB) is proposed. WEDNet is then formed by a cascade of RMFB modules. A batch normalization layer is introduced into the network structure to speed up the network's convergence speed and a dropout method is adopted to suppress the overfitting phenomenon due to the increase of the network’s width and depth. Experimental results show that the method proposed in this paper can reach 99.35% detection accuracy, which can achieve good selection of weak embryos.