LCWG-DETR: a wavelet-enhanced detection transformer for citrus disease detection improving visual perception in...

To address the low detection accuracy and frequent missed detections of citrus diseases caused by leaf occlusion, small lesion sizes, and high visual similarity among disease categories in complex orchard environments, this study proposes LCWG-DETR, a citrus disease detection method that combines wavelet edge features with Gaussian distance regression. First, an LGFE-ResNet18 network is developed for feature extraction. It integrates texture perception guided by the standard deviation of local intensity deviations with global structural encoding in the Fourier frequency domain. This design compensates for the attenuation of high frequency information from small lesions caused by stacked convolutional layers and improves the extraction of subtle disease features. Next, a cross scale adaptive feature fusion module, termed CAFF, is introduced. It employs progressive layerwise fusion and dynamic weighting to reduce the loss of spatial details during upsampling. In addition, a wavelet directional attention module, termed WDAM, is constructed. Haar wavelet decomposition is used to extract horizontal and vertical high frequency components and generate directional modulation weights, thereby guiding the network to focus more accurately on lesion boundaries and regions with abrupt texture variations. Finally, a Gaussian composite distance loss, termed GCD, is introduced. It represents bounding boxes as two dimensional Gaussian distributions with covariance matrices, which alleviates vanishing gradients and scale sensitivity during small lesion regression. Experimental results on the self collected JXDF dataset and the public OFDD dataset show that LCWG-DETR improves mAP by 5.07% and 3.41%, respectively, compared with the baseline model, while reducing the number of parameters by 0.35 M. These results demonstrate that the proposed method achieves a favorable balance between detection accuracy and real time performance. Moreover, it maintains stable detection performance and strong generalization under typical conditions involving occlusion, dense targets, backlighting, and complex backgrounds, providing reliable technical support for the visual perception systems of citrus harvesting robots.