Engineering Vehicle Object Segmentation Algorithm Based on Improved YOLOv8x-seg

To address the challenges of low segmentation accuracy caused by complex backgrounds and relatively small objects in engineering vehicle images, this paper proposes a YOLO-based network, HFF-YOLO, that leverages high-frequency information fusion in the frequency domain. Building on the YOLOv8-seg segmentation network, the High-Frequency Information Extraction module (HFIE) is introduced, transforming images from the spatial domain to the frequency domain. By applying a high-pass filter, the HFIE integrates high-frequency information to achieve precise object edge segmentation in complex backgrounds. Subsequently, a Asymptotic Feature Pyramid Network is employed to construct a shallow segmentation head, and an Adaptive Spatial Feature Fusion module is introduced to address the issue of information loss caused by direct interactions between non-adjacent layers, thereby enabling more effective extraction and fusion of features across different levels. Finally, an engineering vehicle dataset, ENV3K, was constructed by collecting images under power transmission lines, and comparative experiments were conducted. Experimental results demonstrate that HFF-YOLO exhibits significant improvement in small object segmentation compared to YOLOv8x-seg, particularly achieving a 4.7% increase in mAP50 for crane boom segmentation. Additionally, the parameter count of HFF-YOLO is two-thirds that of the original YOLOv8x-seg model.