Enhancing Real-time Detection of Strawberry Diseases: An Optimized SSE-YOLOv5 Model with Improved Accuracy and Small Object Detection

Abstract

Abstract Strawberries are a highly valuable crop widely cultivated across China, making the strawberry industry a crucial component of the country's agricultural sector. Pests and diseases are significant factors affecting the strawberry industry. However, detecting these issues is complicated in real-world environments. Traditional methods, relying on manual inspection by workers, are highly random and inefficient. To achieve real-time detection of strawberry diseases and deploy the model on mobile devices, we propose an improved neural network, SSE-YOLOv5, which enhances detection accuracy while ensuring the model is lightweight. Firstly, we added a channel attention mechanism, SENet, to the original model. The channel attention mechanism offers an effective way to improve feature selection and enhance model performance. It can increase the model's accuracy and recall rate by automatically identifying and emphasizing important features. Additionally, it provides better interpretability, allowing us to see the features that the model focuses on during decision-making. When dealing with data with rich features and complex patterns, the channel attention mechanism shows superior adaptability. Finally, focusing on key features helps to reduce model complexity and the risk of overfitting. Secondly, we added a small object detection layer to improve the accuracy of detecting small targets. Compared with the YOLOv5 baseline model, the improved SSE-YOLOv5 model increased the mAP by 7.4%, reaching 76.3%. The experiments showed that the accuracy of the improved SSE-YOLOv5 model was 75.2%, and the recall rate was 69.8%. The model's detection performance is excellent and can meet the requirements for real-time detection of strawberry diseases.