Applying Multi-layers Feature Fusion in SSD for Detection of Small-Scale Objects

Abstract

AbstractAt present, object detection using convolution neural networks occupies a dominant position. However, the structure of the competent neural network faces hereditary difficulties: high-level networks have large receptive areas and meaning information has high capacity to illustrate, but the solution is diminished, the calculations are weak. In grassroots networks there are relatively small acceptable areas, and it has strong geometric detail information representation capability. SSD (single shot multi-box detection) object prediction of multi-level feature uses mAP, through predicting things with large fields of reception and small fields of prediction of conditions, it also uses high-layer facilities to do. Based on the analysis and introduction of classic SSD algorithms, in this paper, we aim to find small-scale complain at a fast run, through presenting an approach which adapts the SSD detector in relation to trading with precision and speed. An MSSD (Modified single shot multi-box detection) procedure of multistage components combination is offered to supply old knowledge in SSD. The VGGl6 and deep residual networks are used by the MSSD. With Residual network, this article uses the FPN-based network architecture to integrate high and low layers and improves the traditionally sampled structure. The high-layer semantic information is integrated into the low-layer network feature information, and the multi-scale feature mAPs for predicting relapse location case and classification work input are enriched to develop the disclosure efficiency. 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.KeywordsSmall object detectionSingle shot multi-box detectorFeature pyramid networksReal-timeMean average precisionConvolution neural networks