Abstract:
Embryo viability detection is essential for the quality and safety of vaccine production, especially in large-scale manufacturing. Rapid and accurate detection of embryo viability can improve production efficiency and ensure the final quality of vaccines. Traditional machine vision detection methods heavily rely on complex feature extraction algorithms, often designed for specific scenarios. These methods are sensitive to image quality and environmental conditions, meaning that changes in lighting, background, or temperature can affect detection accuracy and stability. Additionally, traditional methods have poor fault tolerance when dealing with noise or abnormal conditions, which limits their applicability in different environments.To address these challenges, this study presents a vaccine embryo viability detection method based on an improved YOLOv8 model, incorporating several innovations to enhance efficiency, accuracy, and adaptability. A specialized image acquisition system was developed to capture high-quality images of embryos incubated for 10 to 11 days, ensuring consistent data across varying environmental conditions. The dataset was expanded using geometric transformations, color adjustments, and image enhancement, increasing the model's robustness in handling diverse image conditions.In terms of model improvements, ShuffleNetV2 was used to replace the YOLOv8 backbone. This change significantly reduces computational complexity while maintaining accuracy, making the model more suitable for deployment on embedded devices where computational power is limited. This replacement enhanced the overall efficiency of the model, supporting its application in large-scale industrial environments.Additionally, a dynamic snake convolutional layer was added to the neck of the YOLOv8 model. This layer can adaptively focus on elongated and curved structures in embryos, making it particularly effective for capturing geometric features of tubular structures. This modification enables the model to more accurately assess the physiological state of the embryos, improving detection precision.Furthermore, the study introduced the EIOU (Embedding Intersection over Union) loss function, designed to handle boundary box alignment and shape similarity more effectively than traditional IOU. EIOU improves the accuracy of boundary box positioning, reducing errors related to the complex shapes of embryos, thereby enhancing the reliability of the model in real-world applications.Experimental results confirmed the superiority of the improved YOLOv8 model in embryo viability detection. The model achieved a precision of 99.2%, a recall of 98.2%, and a mean average precision (mAP50-95) of 96.9%, with increases of 2, 0.3, and 1.5 percentage points, respectively, compared to the original YOLOv8 model. Additionally, the model's computational complexity and inference time were reduced by 60.9% and 60.5%, respectively. These improvements make the model highly suited for large-scale embryo detection applications, providing an efficient, non-destructive method for rapid vaccine embryo viability detection.