ESLC-YOLOv8: Advancing real-time pineapple recognition with lightweight deep learning

Given the current limitations of intelligent pineapple harvesting machinery and the complexity of field environments, significant challenges arise, including the color similarity between pineapples and the background, as well as substantial occlusion and overlap among plants and leaves. This study introduces an enhanced object detection algorithm, EIEStem-v7DS'Sample-LSCD-CA-YOLOv8n (ESLC-YOLOv8), designed for real-time pineapple detection in complex agricultural environments. First, we propose the EIEStem module to enhance the backbone network's convolutional layers, significantly improving edge feature extraction and spatial information preservation. Second, we introduce the v7DS (YOLOv7 DownSample) module to replace conventional downsampling operators, effectively minimizing feature information loss during resolution reduction. Finally, we design a Lightweight Shared Convolutional Detection Head (LSCD) that dramatically reduces model parameters while maintaining detection accuracy, coupled with a Coordinate Attention (CA) mechanism to reinforce critical feature representation. Extensive experimental evaluations indicate that the proposed model achieves the Recall of 0.904 and the mean average precision of 0.945, while reducing the model size to 4.5 MB. Moreover, the number of parameters and floating-point operations decrease by 8.87×10⁵ and 1.6 G, respectively, compared to the original model. The results indicated that the proposed model exhibits superior detection performance for pineapples in complex environments, striking an effective balance between detection accuracy and real-time processing.