In the silkworm breeding industry, silkworm eggs require a green-stimulation treatment before hatching. During this process, it is necessary to daily monitor the developmental stages of the silkworm embryos and adjust the temperature and humidity standards in the green-stimulation room according to the different developmental stages of the embryos to ensure uniform and healthy hatching of the silkworm ants. To achieve intelligent detection of the developmental stages of silkworm embryos, this paper proposes a GB-YOLOv8-based detection model. First, to reduce the model size and the number of parameters, lightweight GhostConv and C2fGhost modules are introduced into the network. Second, the neck network adopts a Bi-directional Feature Pyramid Network (BiFPN) structure, enhancing the feature fusion capability through efficient bi-directional cross-scale connections and weighted feature fusion. Experimental results show that the improved model achieves a mAP value of 95.4% on the self-built dataset, which is an improvement of 2.7% over the original model. Additionally, the model's number of parameters and computational cost are reduced by 0.5M and 0.9G respectively, meeting the practical needs for fast and accurate detection of silkworm embryo development stages.