Multi-scale feature fusion network with FPN and attention mechanism for enhanced object detection

Feature fusion is the process of combining information from different sources or different feature sets into a more comprehensive or informative feature set. Feature fusion is crucial because it can help to synthesize different levels of features to improve detection performance. In order to address the problem of imbalance between semantic and positional features due to the diversity of object scales in real complex scenes, the SSD algorithm is selected as the baseline to study a bi-directional feature pyramid structure based on path enhancement. This structure can fully fuse the semantic information of the deep feature maps and the positional information of the shallow ones. Considering different levels of feature maps have different resolutions and feature information contributes differently to the fusion, a lightweight attention mechanism is utilized for learning the weights of different features, which could effectively fuse features that are inconsistent in semantics and scale. Experimental results show that SSD300 and SSD512 network with feature fusion achieve 79.1% mAP and 81.0% mAP on the test set of PASCAL VOC 2007 dataset, respectively, which are 1.9% and 1.2% higher than the baseline, respectively.