Dual-mode object detection algorithm based on feature enhancement and feature fusion

Abstract In complex conditions, visible light cannot meet the challenge of high-precision target detection requirements, so we propose a dual-modal target detection algorithm, named Dual-Mode Object Detection Based on Feature Enhancement and Feature Fusion (DEF). The algorithm in this paper is improved based on YOLOv5. First, we build a dual-branch feature extraction network for visible light and infrared images, separately extracting features of the images. Subsequently, the extracted feature maps are input separately into a feature enhancement module, which is constituted by a multi-branch convolutional architecture and the ECA (Efficient Channel Attention) mechanism, to acquire an augmented set of salient features. Finally, a module for feature fusion is devised to merge the modality-specific features that are extracted by the two branches. The integrated feature maps are then sent into the subsequent module to achieve dual-mode image target detection. The effectiveness of the proposed algorithm is confirmed through a series of experiments conducted using the FLIR and LLVIP datasets, showcasing its superior performance in practical applications.