Recognizing and detecting the strawberry at multi-stages usingimproved lightweight YOLOv5s

Strawberries have been one of the most popular fruits, due to their the taste and rich nutrition. However, the manual picking cannot fully meet the large-scale cultivation in recent years. Moreover, the short maturity cycle of strawberry can be easy to cause the decay of strawberry fruits, particularly for the untimely picking. Consequently, it is ever increasingly urgent to develop the automatic picking of strawberry. Among them, one of crucial links is the strawberry recognition and detection. The main challenge is to accurately detect the blocked and small target strawberries. In this study, an algorithm was proposed to recognize and detect the obstructed strawberries or the blur target using an improved lightweight Mobile-YOLOv5s. The objective was to enhance both the recognition accuracy and computational speed of strawberry identification during fruit picking. The two-fold strategy was used for the optimization. Firstly, a more lightweight MobileNetV3 network was selected to replace the backbone network of YOLOv5s. The recognition accuracy was improved to reduce the model size and computational complexity. Additionally, Alpha-IOU loss function was also introduced to accelerate the convergence speed of the model. At the same time, the Alpha parameter was used to flexibly adjust the loss function. The recognition accuracy was improved on the blocked and the overlap between strawberries. Secondly, K-Means++ clustering was added to optimize the anchor boxes in target detection. The detection accuracy of blocked strawberries was further optimized to strengthen the detection of small targets. The improved model was also better adapted to the targets with different sizes and shapes. At the same time, the number of detection heads increased to four, in order to expand the receptive field of the model for the detection accuracy of small targets. The experimental results showed that the higher efficacy of the improved model was achieved in the detection frame rate of 44 frames per second (FPS), indicating a significant 15.7% improvement over the original. Furthermore, the computational complexity was measured at 8.3×10⁹ per second, with a substantial 48% reduction from the original. The parameter count was trimmed to 4.5MB with a 15.6% reduction than before. Importantly, the recognition accuracy of mature strawberries reached an impressive 99.5% with an average accuracy of 99.4%, surpassing the original YOLOv5s by 3.6 and 9.2 percentage points, respectively. The more rapid and accurate identification of strawberries was obtained at various stages of growth in practical terms. This technological advancement can hold the promise to realize the intelligent strawberry picking for the high efficiency and productivity in the cultivation of fruits.