Implementation Of Multi-Layers Feature Fusion In SSD For Small Objects

SSD(Single Shot MultiBox Detector)is a popular object detection method.At present,ob-ject detection using convolutional neural networks occupies a dominant position.However,convolutional neural networks have inherent problems in structure:high-level networks have large receptive fields and semantic information has strong ability to represent,but the resolu-tion is low,geometric detail information is weak.The low-level network has relatively small receptive fields,and it has strong geometric detail information representation capability.Al-though the more resolution is high,the more semantic information representation ability is weak.SSD uses multi-scale feature mAPs to predict objects,also uses high-level feature in-formation with large receptive fields to predict large objects,and has low receptive fields for low-level feature information to predict small objects.This brings a problem:When using low-level network feature information to predict small objects,due to the lack of high-level semantic features,SSD have a poor detection effect on small objects.Based on the analysis and introduction of classic SSD algorithms,in this thesis,we aim to detect small objects at a fast speed,through present an approach which adapt the Single Shot Multibox Detector(SSD)with respect to accuracy-vs-speed trade-off as base architecture.A multi-level feature fusion MSSD method is proposed for introducing contextual information in SSD,in order to improve the accuracy for small objects.In detailed,fusion operation are two feature fusion modules is designed,concatenation module and element-sum module,different in the way of adding contextual information.The VGG16 and deep residual net-works are used by the MSSD training to optimize candidate box regression and classification task input feature mAPs to improve detection accuracy and detection speed.With Residual network,this article uses the FPN-based network architecture to integrate high and low lay-ers and improves the traditionally sampled structure.The high-level semantic information is integrated into the low-level network feature information,and the multi-scale feature mAPs for predicting the regression location box and the classification task input are enriched to improve the detection accuracy.Experiments are performed on the logo and VOC2007/2012 datasets which contains a large amount of small objects(objects of 50 pixels or less).Experimental results show that these two fusion modules obtain better mAP on PASCAL VOC2007 and Logo datasets than base-line SSD,especially on some small objects categories.