DF-SSD: An Improved SSD Object Detection Algorithm Based on DenseNet and Feature Fusion

TLDR

The authors propose DF‑SSD, an SSD variant that replaces the VGG‑16 backbone with a DenseNet‑based network and fuses multi‑scale features to improve small‑object detection. DF‑SSD employs a DenseNet‑S‑32‑1 backbone, a multi‑scale feature fusion module, a residual block before prediction, and is trained from scratch. On PASCAL VOC 2007/2012 and MS COCO, DF‑SSD achieves 81.4 %, 79.0 %, and 29.5 % mAP, respectively, improving VOC2007 mAP by 3.1 % over SSD while using half the parameters of SSD and one‑ninth of Faster RCNN, and showing superior performance on small objects.

Abstract

In view of the lack of feature complementarity between the feature layers of Single Shot MultiBox Detector (SSD) and the weak detection ability of SSD for small objects, we propose an improved SSD object detection algorithm based on Dense Convolutional Network (DenseNet) and feature fusion, which is called DF-SSD. On the basis of SSD, we design the feature extraction network DenseNet-S-32-1 with reference to the dense connection of DenseNet, and replace the original backbone network VGG-16 of SSD with DenseNet-S-32-1 to enhance the feature extraction ability of the model. In the part of multi-scale detection, a fusion mechanism of multi-scale feature layers is introduced to organically combine low-level visual features and high-level semantic features in the network structure. Finally, a residual block is established before the object prediction to further improve the model performance. We train the DF-SSD model from scratch. The experimental results show that our model DF-SSD with 300 × 300 input achieves 81.4% mAP, 79.0% mAP, and 29.5% mAP on PASCAL VOC 2007, VOC 2012, and MS COCO datasets, respectively. Compared with SSD, the detection accuracy of DF-SSD on VOC 2007 is improved by 3.1% mAP. DF-SSD requires only 1/2 parameters to SSD and 1/9 parameters to Faster RCNN. We inject more semantic information into DF-SSD, which makes it have advanced detection effect on small objects and objects with specific relationships.