Answers are grounded in this paper and cite their sources.
• The present study puts forward a Contextual Guidance Module (CDB) as a means to address target size variation in pest detection by integrating dynamic convolution, hierarchical receptive fields, and global-local feature collaboration. • A novel DFANet structure with two custom modules has been designed to enhance multi-scale context extraction and frequency-aware feature representation. This design facilitates the efficient decoupling and fusion of high- and low-frequency features, thereby enhancing the model's adaptability in complex detection scenarios. • It is posited that an enhanced version of the C2f structure, designated C2fDFA, is proposed to enhance multi-scale and contextual feature extraction. This proposition is grounded in the principles of DFANet. Furthermore, a field-collected dataset containing ten crop insect species has been constructed to support intelligent pest detection in agriculture. In the field of smart agriculture, vision-based pest detection is crucial for effectively monitoring and managing pest populations. However, existing methods often show low accuracy in dense scenes, particularly when dealing with multiple pest categories and small or minuscule targets. Additionally, the large and complex backbone networks impose limitations on deploying these models on monitoring devices with limited resources. To address these challenges, a lightweight, context-aware, and dynamic feature adaptive pest detection network (CDFA-Net) is proposed. CDFA-Net designs a Context-Driven Block (CDB) to effectively capture richer contextual information, thereby mitigating accuracy loss caused by occlusions. Furthermore, a Dynamic Frequency Adaptive Network (DFANet) and a new feature extraction module, C2fDFA, have been developed to enhance feature retention during extraction and improve the detection of targets affected by poor localization or interference from large objects. The detection head has been optimized to improve performance without compromising the lightweight design. Comprehensive evaluations on the self-constructed CP10 dataset and the public YST dataset demonstrate the effectiveness and efficiency of the proposed method. On the CP10 dataset, CDFA-Net achieved an average detection accuracy and F1 score of 82.9% and 82%, respectively, surpassing the baseline by 8.1% and 6%, with a parameter reduction of 2.79MB, equivalent to a decrease of 0.22MB. On the YST dataset, CDFA-Net achieved mAP and F1-Score of 91.6% and 90.6%, respectively, demonstrating a 7.9% and 7.7% improvement over the baseline. A comparison of CDFA-Net's performance with other existing algorithms demonstrated that CDFA-Net exhibited the optimal overall performance in the context of multi-object detection.
No takes yet. Share an insight, caveat, or question.
Le et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: