Insect pests on yellow sticky traps are notoriously difficult to detect automatically because each specimen occupies only a few dozen pixels and is often surrounded by a vast, feature-poor background. This study proposes a fully automated pipeline that combines (i) HSV-based trap isolation, (ii) self-bootstrapped pseudo-labelling with YOLOv8 and YOLOv11, and (iii) a slicing algorithm that converts every full-resolution frame into non-overlapping 640 × 640 patches, each containing a set of un-truncated insects. On the revised Yellow Sticky Traps dataset (8,114 annotations, three classes), this slice-based training strategy markedly improves performance. For YOLOv8, mAP@[0.5:0.95] rises from 0.41 to 0.77, with precision/recall increasing from 0.82/0.77 to 0.92/0.93. For YOLOv11, mAP@[0.5:0.95] increases from 0.46 to 0.89 (a 92% relative gain), while precision and recall reach 0.96. Confusion matrix analysis confirms that slice-based models virtually eliminate background bleed-through and inter-species misclassification, problems that severely affect full-image baselines. The pipeline requires no manual annotation beyond an initial calibration phase, making it well-suited for large-scale, real-time pest surveillance in commercial greenhouses and open-field agriculture.