Improvements to Deep Learning-Based Object Detection Algorithms in Greenhouses

Capsicums (chili peppers), serving as critical economic crops, hold a significant role within greenhouse cultivation, driving the development of the agricultural economy. Traditional methods for detecting chili pepper flowers are notably time-consuming and labor-intensive. The unique visual characteristics of chili pepper flowers are suitable for accurate classification and identification using deep learning technology. Nevertheless, numerous challenges exist in identifying chili pepper flowers in complex backgrounds. This study proposes an edge-enhanced Normalization-based Attention Module (E-NAM), which integrates gradient and magnitude into NAM. E-NAM is embedded into the C2f (Coarse-to-Fine) module of YOLOv8 to enhance edge-awareness and feature discrimination. To enhance the model’s semantic understanding, we propose ES-SPPF (SPPF with enhanced semantic transmission), making it more suitable for small object detection. At the same time, a threshold-based piecewise IoU (PIIoU) is proposed to improve the model’s inference efficiency. Finally, dilated convolution is introduced between the backbone and the head, which helps to capture small objects by integrating local and global information. Compared with the vanilla YOLOv8, our proposed enhancements (NAM, ES-SPPF, PIIoU, and Dilated Convolution) achieve a consistent and significant improvement of 4.4% in mAP (Mean Average Precision) (from 78.2% to 82.6%) on the greenhouse chili flower dataset, while maintaining a real-time inference speed of 73 FPS (Frames Per Second). The findings can serve as an important reference for crop object detection.