Multi-crop leaf disease detection (MCLDD) is crucial for smart agriculture. This study proposes a lightweight disease detection framework, MMW-YOLO, to address several challenges in MCLDD under field conditions, including inconspicuous symptoms, strong background interference, and large variations in lesion scales. First, the wavelet transform and SS2D are incorporated into C3k2 to build a feature enhancement module that improves feature representation in complex backgrounds. Second, a subband-aware wavelet downsampling module is proposed that integrates anti-aliasing pre-filtering, Haar wavelet decomposition, and subband attention recalibration to reduce the loss of fine-grained details during scale transformation. Finally, a multi-scale adaptive feature pyramid is designed, with P2-assisted feature injection and secondary refinement introduced at the P3 stage. Combined with a multi-branch feedback mechanism, this design further enhances multi-scale feature fusion. Experiments on a self-constructed dataset containing 15 classes from five crop species show that MMW-YOLO improves mAP@50 and F1 score by 2.63\% and 3.06\%, respectively, compared with the baseline model, while reducing the number of parameters to 2.42 M. These results indicate that the proposed method achieves a favorable balance between detection accuracy and deployment efficiency. The data and code will be openly available in GitHub upon publication.