Robust real-time detection of small objects under occlusion and appearance similarity in orchard environments: A case study on banana pseudostems

In banana plantations, pseudostems are slender cylindrical structures crucial for plant support and nutrient transport. Accurate detection is vital for growth monitoring, yield estimation, and automation. Compared with conventional small object detection tasks under occlusion and complex backgrounds, banana plantations introduce unique challenges: pseudostems appear as slender, densely packed small targets, are often occluded by banana leaves, and are visually similar to branches and bamboo poles. These lead to severe occlusion, appearance-similar interference. To address these issues, we proposes a real-time detection method deployable on edge devices with enhanced fine-grained small object perception and improved robustness to occlusion and background interference. First, to mitigate the issue of feature loss for slender and small-scale pseudostems during downsampling, an SPD-Conv module is introduced to preserve local structural details, while the proposed DynaFusion Neck further enhances slender object representation via adaptive upsampling and key-region focus. Second, to handle severe occlusion by banana leaves and significant scale variation across pseudostem growth stages, DynaFusion Neck incorporates multi-layer contextual features and a newly added high-resolution detection head, enhancing multi-scale alignment for robustness in detecting occluded small targets. Finally, a Bayesian Dynamic Threshold Loss (BDTL) dynamically adjusts decision boundaries based on sample uncertainty, enhancing robustness against appearance-similar background interference. Experimental results demonstrate that the proposed method achieves excellent performance in banana pseudostem detection and is well-suited for real-time, high-precision deployment on edge robotic devices. • A real-time detection framework is proposed for robust detection of slender small targets under severe occlusion and visually similar interference in complex orchard environment. • A multi-scale fusion neck is proposed to enhance feature alignment under scale variation and occlusion, and jointly improve detection accuracy with a fine-grained convolution module. • A Bayesian Dynamic Threshold Loss is designed to dynamically adjust sample weights and suppress appearance-Similar Interference in semantically similar and complex backgrounds. • The proposed model achieves higher accuracy and recall with fewer parameters, and has been practically deployed on edge devices for orchard robots.