RGB Image-Based Lightweight Eggplant Fruit Disease Recognition via Frequency Spatial Collaborative Downsampling and Dynamic Attention

Eggplant is an important vegetable crop, and accurate identification of fruit diseases is critical for ensuring yield and quality. However, field-captured eggplant images may contain small disease lesions, foliage occlusion, and illumination variations, posing challenges for accurate disease recognition. To address these challenges, this paper proposes E-GhostNetV3 (Eggplant-GhostNetV3), a lightweight disease recognition network for field-captured images. The core improvements include a frequency–spatial collaborative downsampling module (WCD-Down), which combines Haar-based multiband decomposition with parallel convolution for spatial feature extraction, and a Dynamic Spatial–Channel Attention (DySCA) module to improve spatial and channel-wise feature representation. Evaluated on an eggplant disease dataset comprising 1823 real field-captured images across five classes, E-GhostNetV3 achieves 89.61% Top-1 accuracy with 1.81 M parameters and an estimated 116.60 M MACs under the default THOP counting rules, outperforming the GhostNetV3 baseline (87.07%) by 2.54 percentage points. These results demonstrate the effectiveness of E-GhostNetV3 on the evaluated dataset and its potential as a lightweight model for edge-based smart agriculture applications. The study also provides a benchmark for evaluating models trained from scratch under data-scarce conditions.