Mifanet: multi-scale information fusion attention network for determining hatching eggs activity via detecting PPG signals

It is crucial to accurately detect dead embryos when developing vaccines. Detection of photoplethysmography signals can help determine hatching egg activity via convolutional neural networks, which is considered the most reliable and promising method in the industry. The existing detection methods face some challenges because the convolution operations have difficulty capturing global representations from PPG signals. In this study, we propose a composite network structure, termed MifaNet, to take advantage of self-attention mechanisms for enhanced convolution operations. MifaNet captures long-distance feature dependencies and local feature details through independent branches. A feature fusion module is used to fuse concurrent features and output the detection results. Extensive experiments were conducted on our dataset, and the results confirm that MifaNet outperforms state-of-the-art techniques.