Object Detector with Multi-head Self-attention and Multi-scale Fusion

Abstract

Aiming at the problems that the existing object detection network cannot fully extract the internal correlation of features and the multi-scale feature fusion is insufficient, an object detector that introduces multi-head self-attention and multi-scale fusion is proposed. Firstly, the multi-head self-attention mechanism is introduced, and the internal correlation of the extracted features is strengthened through the multi-head self-attention mechanism, and the dependence of the features on the external information is reduced. Secondly, the DenseASPP module is integrated in the designed network, and the ability of multi-scale feature fusion is improved through dilated convolution with different dilatation rate. The experiments of the proposed object detection algorithm on the PASCAL VOC 2007 dataset show that the overall accuracy is greatly improved compared with other advanced algorithms.