ADSSD: Improved Single-Shot Detector with Attention Mechanism and Dilated Convolution

Abstract

The detection of small objects is easily affected by background information, and a lack of context information makes detection difficult. Therefore, small object detection has become an extremely challenging task. Based on the above problems, we proposed a Single-Shot MultiBox Detector with an attention mechanism and dilated convolution (ADSSD). In the attention module, we strengthened the connection between information in space and channels while using cross-layer connections to accelerate training. In the multi-branch dilated convolution module, we combined three expansion convolutions with different dilated ratios to obtain multi-scale context information and used hierarchical feature fusion to reduce the gridding effect. The results show that on PASCAL VOC2007 and VOC2012 datasets, our 300 × 300 input ADSSD model reaches 78.4% mAP and 76.1% mAP. The results outperform those of SSD and other advanced detectors; the effect of some small object detection is significantly improved. Moreover, the performance of the ADSSD in object detection affected by factors such as dense occlusion is better than that of the traditional SSD.