LRSFT: UAV wheat image super-resolution via a lightweight recursive spatial frequency-domain transformer with wavelet loss
摘要
Unmanned aerial vehicle (UAV) equipped with visible light sensors offer a low-cost and efficient solution for monitoring wheat growth and disease. However, limitations from sensor resolution and flight altitude often result in low-quality imagery. Existing super-resolution (SR) methods for UAV-acquired wheat images often suffer from low feature utilization, insufficient frequency-domain representation, and high computational cost. To address this, we propose a lightweight recursive spatial frequency-domain transformer (LRSFT) with wavelet loss, which can enhance high-frequency detail recovery while maintaining low computational overhead. First, we design a recursive spatial module (RSM) to iteratively aggregate local spatial features with an expanding effective receptive field and model contour structural relationships. This generates a spatial recursive map while significantly reducing computational overhead. Second, a frequency-domain enhancement module (FEM) is employed to capture high-frequency features via a global receptive field. These spatial and frequency-domain maps are fused through residual connections to form a spatial-frequency feature map, enhancing reconstruction performance. Finally, we introduce a wavelet loss to address the insufficient learning of high-frequency information when using