Xiaomila Green Pepper Target Detection Method under Complex Environment Based on Improved YOLOv5s

Real-time detection of fruit targets is a key technology the Xiaomila green pepper (Capsicum frutescens L.) picking robot. The complex conditions orchards make it difficult to achieve accurate detection. However, most existing deep learning network algorithms cannot effectively detect fruits covered by leaves, branches, and other in natural scenes. As detailed this paper, Red, Green, Blue (RGB) images mature stage were collected under light for building dataset an improved YOLOv5s model (YOLOv5s-CFL) proposed improve efficiency adaptability robots environment. First, convolutional layer Cross Stage Partial (CSP) replaced with GhostConv, speed through lightweight structure, accuracy enhanced adding Coordinate Attention (CA) replacing Path Aggregation Network (PANet) neck Bidirectional Feature Pyramid (BiFPN). In experiment, YOLOv5s-CFL was used Xiaomila, results analyzed compared those original YOLOv5s, YOLOv4-tiny, YOLOv3-tiny models. With these improvements, Mean Average Precision (mAP) 1.1%, 6.8%, 8.9% higher than YOLOv3-tiny, respectively. Compared YOLOv5 model, size reduced from 14.4 MB 13.8 MB, running 15.8 13.9 Gflops. experimental indicate that improves has good real-time performance application prospects field robots.