YOLO26-P2-P4-SFAD: an improved small object detection method based on P2-P4 restructuring and wavelet-attention frequency domain downsampling
Abstract
Small object detection in UAV aerial imagery and debris detection scenarios faces persistent challenges: limited pixel occupancy, weak texture information, and inevitable detail loss during downsampling. This paper presents a structural improvement to the YOLO26 detection framework tailored for small object detection. First, we adopt the official YOLO26-P2 as the baseline, leveraging its P2 high-resolution detection branch to enhance small-object representation. We further remove the P5 detection head to construct YOLO26-P2-P4, concentrating computational resources on high-resolution feature layers. Building upon this, we design the SFAD (Spatial-Frequency Attentional Downsampling) module, which employs a dual-branch architecture: a spatial convolution branch for local structure modeling and a Haar wavelet frequency branch for high-frequency detail preservation, followed by SimAM parameter-free attention for salient region recalibration.Comparative experiments with representative downsampling schemes, including AConv, ADown, HWD, and SCDown, are conducted. On the Debris Detection.v43i.yolo26 test set, YOLO26-P2-P4 improves mAP@0.5 from 39.67% to 45.43% and mAP@0.5:0.95 from 18.41% to 20.70% compared with YOLO26-P2. Incorporating SFAD further boosts mAP@0.5 to 46.55% and mAP@0.5:0.95 to 21.21%, with AP_s reaching 17.76%. On the VisDrone2019-DET test set, YOLO26-P2-P4-SFAD achieves 25.53% mAP@0.5, 14.35% mAP@0.5:0.95, and 6.89% AP_s, consistently outperforming competing baselines. The proposed method effectively enhances small-object responses with minimal parameter increase, demonstrating strong research value and engineering potential.