Enhanced Small Object Detection Neural Network

Abstract

Faster-RCNN is an vital deep learning object detection algorithm. Nevertheless, the small object detection effect of Faster-RCNN, which does not use multi-layer feature map, is not good enough. In this paper, a new network architecture called enhanced small object detection neural network (ESOD) is proposed. Two advancements are proposed in ESOD. Firstly, a innovative multilayer combining network which integrating multi-layer feature map is designed to leverage multi-layer convolutional information. Thereupon, the ability of small object detection can be improved. Secondly, the scale-transfer layer (STL) is introduced in the fusion of features which are in high layer and features which are in low layer, which extends the size of the features by reducing the quantity of aisles. In this way, the semantic features of distinct size are acquired without additional parameters. Extensive experimental results on PASCAL VOC 2007 and 2012 datasets indicate that, compared with several commonly used algorithms, ESOD brings better small object detection ability than existing methods for object detection.