Robust detection for selective harvesting of field flat jujube: overcoming occlusion and small-target challenges in...

The detection of field flat jujube is constrained by their characteristics and the complex agricultural environment, presenting challenges such as small target size and dense occlusion, which can easily lead to insufficient information in occluded areas and the loss of features in small targets. This paper builds upon the latest YOLOv12 (You Only Look Once) network, focusing on compensating for information loss in occluded regions and improving detection accuracy for small objects. Firstly, the Separated and Enhancement Attention Module (SEAM) was incorporated into the neck network to enhance feature representation in occluded regions. Secondly, to enrich contextual semantic information in dense detection tasks, we replaced the up-sampling operator with the Content-Aware Reassembly of Features (CARAFE) operator. Finally, the Parallelized Patch-aware Attention (PPA) module was integrated into the detection head to design a small-target-specific detection head with a built-in attention mechanism, through which the interactive fusion of global and local feature representations was realized. Experimental results demonstrated that the proposed YOLOv12-SCP network achieved a mean average precision (mAP@0.5) of 96.8% and an F1-score of 0.91, surpassing the original YOLOv12n model by 1.2% and 1.0%, respectively. Meanwhile, the mAP@0.5:0.95 increased by 2.7 percentage points, with the average precision for ripe and unripe fruits reaching 97.6% and 96%, respectively. Through extensive ablation experiments and comparisons with current mainstream object detection methods, it is demonstrated that this method exhibits superior performance in detecting small object occlusions in complex environments.