A Real-Time Apple Maturity Detection Method Combining Lightweight Networks and Multiscale Attention Mechanisms

Apple maturity detection algorithms based on deep learning typically involve a large number of parameters, resulting in high computational costs, long processing times, and dependence on high computational power graphics processing units (GPUs). This paper proposes an improved YOLOv8 model to address the issues related to the maturity detection of Fuji apples grown in China using image-based methods. The model was optimized in several ways according to the characteristics of apple targets and scenes. First, a lightweight MobileNetV3 is used as the backbone network, replacing the original CSPDarknet-53 backbone network, which reduces the model parameters and computational complexity and increases the inference speed. Second, by introducing the efficient multiscale attention (EMA) module and using the bidirectional feature pyramid network (BiFPN) in the neck part, the model enhances the extraction capability of important features and suppresses redundant features, thus improving the model’s generalization ability. Experimental results show that the size of the model is 2.6 megabytes. On the apple dataset, its precision, recall, F1 score, and mean average precision reach 90.2%, 88.5%, 89.3%, and 91.3%, respectively, with improvements of 4.3%, 3.2%, 3.7%, and 2.6% compared to the original model. Based on this model, an Android application has been developed for real-time apple maturity detection. The improved model proposed in this paper achieves real-time apple target recognition and maturity detection, providing quick and accurate target recognition guidance for the mechanical automatic harvesting of apples.