YOLO-FWF: a lightweight frequency–wavelet feature fusion network for robust scene text detection

Scene text detection remains challenging due to large scale variations, slender stroke structures, complex backgrounds, low contrast, and arbitrary text orientations. To address these issues, this paper proposes YOLO-FWF, an improved lightweight scene text detection model based on YOLO11n. First, a C3k2_FrequencyCM module is introduced into the backbone to enhance fine-grained text representation by combining Fourier-domain amplitude compensation with spatial high-frequency feature enhancement, thereby improving the sensitivity of the network to text strokes, edges, and weak texture cues. Second, a C2PSA_WDAM module is designed to replace the original C2PSA structure. By decomposing features into low-frequency semantic components and high-frequency detail components through wavelet transform, the module uses high-frequency information to guide attention modeling and suppress background interference. Third, a Fourier Angle Alignment Fusion module is proposed for multi-scale feature fusion. It estimates dominant local orientations in the frequency domain and aligns high-level semantic features with low-level structural features before fusion, improving the detection of inclined, curved, and irregular text instances. Experiments on two Roboflow-hosted scene text datasets demonstrate that YOLO-FWF achieves superior detection accuracy while maintaining a lightweight structure. On the text scene computer vision dataset, YOLO-FWF reaches 86.8% mAP50 and 64.2% mAP50:95 with only 2.89 M parameters and 6.7 GFLOPs. On the text detection dataset, it achieves 54.7% mAP50 and 34.6% mAP50:95, outperforming YOLO11n and other mainstream detectors under the same evaluation setting. These results verify the effectiveness of frequency-domain enhancement, wavelet-guided attention, and direction-aware feature fusion on the evaluated datasets. Because the current experiments do not include standard public scene-text benchmarks, broader generalization remains to be further verified.