YOLOv11n-SWAQ: A Lightweight Adaptive Compression Method for Agricultural Images

Agricultural pest and disease images often lose diagnostically important textures during transmission and storage, while conventional codecs struggle to balance high compression ratios and information fidelity. This study proposes YOLOv11n-SWAQ, a lightweight adaptive compression framework integrating shared feature encoding and critical-region-aware quantization. The YOLOv11n backbone and feature fusion network are reused as a shared encoder to reduce redundant feature extraction. A lightweight window-based attention module enhances lesion boundaries, insect contours, and local textures. Detection boxes and attention weights are combined to construct a three-level quantization map for background, ordinary critical, and high-importance texture regions, while a hyperprior entropy model improves latent probability estimation. At an actual complete-stream bitrate of approximately 0.31 bpp, the proposed method achieved an ROI-PSNR of 36.970 dB and an ROI-MS-SSIM of 0.9652. Compared with the YOLOv11n+CAE cascaded baseline, the reconstruction error in complex-texture regions decreased by 4.8%, indicating preferential bitrate allocation to diagnostically important regions. The proposed method achieved an end-to-end processing time of 13.44 ms, representing an average reduction of approximately 72% compared with representative advanced learned image compression models. Overall, YOLOv11n-SWAQ provides a favorable trade-off between critical-region fidelity and computational efficiency for high-ratio agricultural image compression, with its principal advantage lying in lightweight task-oriented processing rather than a large absolute gain in global reconstruction quality.