# DFSNet: directional feature aggregation and shape-aware supervision for eggplant pest and disease detection

## TL;DR

DFSNet is a new model for detecting pests and diseases on eggplants in natural settings, improving accuracy and speed for small and complex targets.

## Contribution

DFSNet introduces directional feature aggregation and shape-aware supervision for improved detection of eggplant pests and diseases.

## Key findings

- DFSNet achieves 81.0% precision, 78.3% recall, and 80.5% mAP@50 on eggplant pest and disease detection.
- The model improves small target detection accuracy and robustness in complex backgrounds.
- DFSNet has 1.8M parameters and 5.4G computational cost with an inference speed of 378.13 FPS.

## Abstract

In natural planting environments, pest and disease detection on eggplant fruits is characterized by small lesion sizes, weak edge feature information, significant scale variations, and complex backgrounds. Particularly, fruit borer holes, fruit rot lesions, and melon thrips bite marks exhibit obvious differences in size, edge structure, and spatial distribution, posing considerable challenges for real-time accurate detection. This paper proposes the DFSNet, a lightweight improved network for pest and disease detection on eggplant fruits in natural scenes. First, PConv is introduced in the P1, P2 shallow feature extraction stages of the baseline model’s backbone network to enhance the modeling capability for fine-grained directional textures and weak edge information. Subsequently, an MSDA (Multi-Scale Directional Aggregation) module is designed and embedded into the feature enhancement modules at the P3, P4, and P5 layers of the backbone, which effectively improves the perception capability for insect hole edges and lesion contours through multi-directional depthwise separable convolution and Directional Edge Enhancer (DEE). Furthermore, a CSP-MSLA structure is introduced into the neck network, combining multi-scale linear attention mechanism with cross-stage partial connections to achieve selective enhancement of key pest and disease regions while maintaining low computational complexity. Finally, an SDDH (Shape-based Dynamic Detection Head) is introduced, which enhances the model’s adaptive capability to different pest and disease geometric features and scale variations by introducing Scale-based Dynamic Loss. Experimental results demonstrate that the model achieves Precision of 81.0%, Recall of 78.3%, and mAP@50 of 80.5% on a self-constructed eggplant pest and disease dataset under natural scenes, representing improvements of 6.9, 8.8%, and 7.8% percentage points respectively compared to the baseline model. Meanwhile, the model parameters and computational cost are compressed to 1.8M and 5.4G respectively, with an inference speed of up to 378.13 FPS. The proposed method effectively improves small target detection accuracy and robustness under complex backgrounds while ensuring real-time performance, demonstrating particularly significant advantages in detecting small targets such as fruit borer holes and melon thrips bite marks, proving that this model is an efficient and robust real-time detection model for eggplant fruit pests and diseases.