Enhancing object detection with large kernel convolution and cross convolution

doi:10.1016/j.dsp.2025.105433

Return

Enhancing object detection with large kernel convolution and cross convolution

2025-06-25

PRE

DOI：10.1016/j.dsp.2025.105433

Original

Original request for help

Save

Abstract

En 中文

Existing object detection models often struggle with detecting small objects due to their limited ability to capture sufficient contextual information. In this paper, we introduce a lightweight object detection model that leverages large kernel convolution with attention (LKA) and a hierarchical feature fusion group (HFFG) to address this issue. The LKA module employs large kernel convolution to capture long-range dependencies and contextual information, combined with depthwise separate convolution to maintain a lightweight design. An incorporated attention mechanism further enables the modal to adaptively focus on key areas, thereby improving detection performance for small objects. The HFFG module, which integrates Cross Convolution Blocks, explores and retains structural information across different scales. By effectively extracting structural details, our model exhibits enhanced performance on object of various sizes. Extensive experiments on the VisDrone2019 and PASACAL VOC datasets demonstrate that our model achieves an outstanding mAP of 23.4 %, surpassing the baseline YOLOX-s model by +1.5 %. These results not only validate the effectiveness but also demonstrate its robustness and generalization capability.