Abstract

Apple leaf disease detection in orchards faces a unique challenge: lesion regions (foreground) are semantically important but visually less salient (weak texture, blurred boundaries), while background distractors (soil, weeds) are visually salient but semantically irrelevant. Traditional models, reliant on visual saliency, are usually misled by strong background textures. Unlike humans who first use leaf shape to locate possible lesions and then examine local textures, these models lack this semantic-guided strategy, leading to inaccurate detection. To address this problem, we propose a Semantic-Decoupled Dual Attention Network (SD-DAN). The core idea is to achieve target-distractor separation via semantic decoupling and then foreground enhancement and background suppression via dual attention. The network consists of two core modules. First, a frequency-enhanced preprocessing module adaptively fuses low- and high-frequency information to highlight lesion details and suppress redundant low-frequency information. Second, an attention-based foreground-background decoupling and enhancement module explicitly separates foreground and background with a learnable threshold, and applies a dual-attention mechanism: spatial-frequency enhancement to the foreground (active focusing) and spatial-frequency suppression to the background (active filtering), to improve discriminative features of apple disease. Experiments on the Apple Leaf Disease Dataset 9 (ALDD9) and the Apple Leaf Diseases Dataset demonstrate that SD-DAN achieves strong performance and consistently outperforms mainstream detectors across nine common apple disease detection tasks.

IPC Classification

Keywords