A lightweight detection model for greenhouse-cultivated strawberries based on YOLOv5

Abstract

The efficient detection of strawberries has great significance in the realization of strawberry production estimation and automatic picking in the greenhouse. Factors such as the complex growing environment in the field and fruit aggregation shading, especially for immature strawberries at the turning stage and mature strawberries with high similarity in shape, size, and even color, resulted in low accuracy of detection. This research aims at establishing an improved lightweight model based on YOLOv5 which is for strawberry ripeness detection in natural environments, in response to the problems of some current detection models whose structures are complex and difficult to deploy on low-cost devices. The method first adds a smaller target detection layer to the original network, then removes part of the deep structure of the network to reduce the complexity of the model, by adjusting the depth-multiple and width-multiple partially to achieve a lighter model while ensuring accuracy. After that, the performance of the model was further improved by introducing the BiFPN structure and SimAM attention module. The experimental results show that the improved model provides an effective method for detecting strawberries in natural environments. Compared with the YOLOv5s model, the improved model has a 60.35% reduction in parameters and a 55.47% reduction in the model size, and the mAP, mAP0.5:0.95, and F1 are improved to 91.86%, 79.04%, and 90.9%, respectively. The model proposed in this research proves an excellent generalization performance in three strawberry test sets and provides a reference for subsequent deployment in small/low-cost picking robots.