A Multi-Scale Edge Feature Network for Robust Object Detection

Abstract

Object detection, a foundational task in computer vision, entails accurately identifying and localizing objects in images, which remains challenging due to issues like object occlusion and multiscale detection imbalance. This paper proposes the Multi-Scale Edge Feature Enhancement Network (MEFENet), a novel one-stage object detection framework designed to address these challenges. MEFENet introduces two key innovations: (1) the Multi-Scale Edge Feature Extraction (MEFE) structure, which fuses extracted edge features with multi-scale feature maps, enriching semantic representations to improve occluded object detection; and (2) the Receptive Field Enhancement (RFE) module, which refines feature semantics and mitigates multiscale detection imbalances. MEFENet leverages a residual network (ResNet) backbone and combines outputs from the Feature Pyramid Network (FPN) and MEFE structures, which are subsequently processed through the RFE module for enhanced semantic feature extraction. Extensive experiments on the PASCAL VOC 2007+2012 and Microsoft COCO datasets demonstrate that MEFENet achieves state-of-the-art detection accuracy, outperforming nine representative methods in key evaluation metrics. These results validate the effectiveness of the proposed innovations in addressing occlusion and multiscale detection challenges.