SLP-Net: A Dual-Level Contrastive Learning Framework with Stripe Attention for Elongated Pepper Detection in Complex Field Environments

Pepper detection in field images is difficult because the fruits can differ substantially in appearance, and many are partially covered by nearby leaves. Localization becomes less reliable when a pepper is slender or when only part of its contour is visible. SLP-Net was developed for this setting. Rather than increasing model size, it is designed to preserve shape cues that are easily weakened in cluttered field scenes. This makes the detector less sensitive to differences among pepper instances and to cases in which the visible region is incomplete. On PP-Set, SLP-Net outperforms the compared detectors, with clearer gains at higher IoU thresholds and on small targets. A similar pattern is observed on CH-Set, where disease, deformation, and stronger background interference further increase the difficulty of detection. Overall, these results indicate that SLP-Net remains more stable when pepper targets vary more strongly in geometry, surface condition, and visibility.