FBENet: A Highway Road Debris Detection Network Based on Frequency-Aware Bidirectional Feature Fusion and Efficient Attention Enhancement

Highway road debris is often small, irregular, weakly textured, and poorly contrasted against complex backgrounds, while limited real-world samples constrain model generalization. This study constructs a Synthetic-Debris dataset by compositing three representative debris categories—brick, paper box, and rock—onto highway scenes to support small-object and low-contrast detection. FBENet (Frequency-Aware Bidirectional Feature Fusion and Efficient Attention Enhancement Network) is developed from YOLO11n using a task-oriented, stage-coupled design: FreqFusion (Frequency-aware Feature Fusion) is embedded at two top-down cross-scale fusion stages to preserve boundary details and improve cross-resolution consistency before learnable bidirectional aggregation by BiFPN (Bi-directional Feature Pyramid Network), while EMA (Efficient Multi-Scale Attention) is applied only to the final high-resolution P3 feature before detection. The contribution lies in the stage-specific organization of established operations rather than in proposing new primitive modules. On Synthetic-Debris, FBENet achieved an mAP@0.5 of 0.862, an mAP@0.5:0.95 of 0.664, and an F1-score of 0.811, with 2.43 M parameters and 6.8 GFLOPs. In the fixed synthetic-to-real split, FBENet exceeded YOLO11n by 5.0 and 2.8 percentage points in mAP@0.5 and mAP@0.5:0.95, respectively. Five-fold cross-validation showed modest mean AP gains and lower overall AP variance, although transfer behavior remained class-dependent. Runtime evaluation showed that FBENet incurred additional latency relative to YOLO11n. Under PyTorch FP32, the network inference latency increased from 8.07 to 12.79 ms, while the end-to-end throughput decreased from 79.15 to 53.97 FPS. Under TensorRT FP16, FBENet achieved 126.59 FPS compared with 137.64 FPS for YOLO11n. These results indicate that FBENet improves detection accuracy at the cost of additional runtime, representing an accuracy–latency trade-off rather than an improvement in inference efficiency. Overall, frequency-aware bidirectional fusion and selective high-resolution attention are useful for improving highway debris detection under limited real-data conditions.