YOLO-MFE: towards more accurate object detection using multiscale feature extraction

Authors:

Abstract
YOLOv3, as a multi-scale object detection algorithm, has a simple structure and can be detected quickly. However, during the training process, as the bottom convolutional layer contains more object detail, and the information in the underlying layer will be gradually faded out through continuous convolution. To address this, this paper adds a normalized attention mechanism behind the feature layers at different scales to further suppress insignificant features and enhance the discriminability and robustness of the features. Moreover, an improved pyramid structure is leveraged to fuse the detail of the shallow feature maps with the semantic feature information of the higher layers, aiming to improve the model performance. In particular, a YOLOv3 object detection algorithm named YOLO-MFE based on improved multi-scale feature extraction is proposed. Experimental results on the Pascal VOC dataset demonstrate that our algorithm outperforms the original YOLOv3 algorithm in terms of object detection accuracy, with a mAP on the Pascal VOC 2007 test set that is 1.66% higher.

This publication has 4 references indexed in Scilit: