Detecting Green Pepper Fruit and Peduncles Based on Improved YOLOv8n

Accurately and efficiently identifying green pepper fruits and peduncles is crucial for achieving automation in green pepper harvesting. To address the challenges posed by complex background interference and branch and leaf occlusion in greenhouse environments, we propose an enhanced model, YOLO-CEDM, based on YOLOv8n. This model introduces a Wavelet Transform-Based Pooling method to form the Conv-BatchNorm-SiLU-WaveletPool (CBS-W) module, improving small object detection accuracy and reducing aliasing effects. Additionally, we integrated an Efficient Channel Attention (ECA) mechanism onto the SPPF layer of YOLOv8n to further enhance its performance. DySample, a dynamic adaptive sampling method, was employed to replace the up-sampling module, enhancing the model's capacity to extract features. Lastly, we incorporated the Minimum Point Distance IoU (MPDIoU) to strengthen the model's robustness against aspect ratio and scale variations and improve its sensitivity to overlapping regions. The experimental results demonstrate that YOLO-CEDM achieved an average detection accuracy of 92.3% with an average detection time of 13.2 ms, along with 3.02M parameters, 11.9G FLOPs, and a 9.5MB model size. The model's average detection accuracy surpassed traditional neural networks by 7.5% to 15.8%. As a result, our model satisfies the demands for lightweight design and real-time performance, offering a theoretical foundation for the automated harvesting of green pepper fruits.