Abstract

In this paper, a parallel SSD (Single Shot MultiBox Detector) fusion network based on inverted residual structure (IR-PSN) is proposed to solve the problems of the lack of extracted feature information and the unsatisfactory effect of small object detection by deep learning. Firstly, the Inverted Residual Structure (IR) is adopted into the SSD network to replace the pooling layer. The improved SSD network is called deep network of IR-PSN to extract high-level feature information of the image. Secondly, a shallow network based on the inverted residual structure is constructed to extract low-level feature information of the image. Finally, the shallow network is fused with the deep network to avoid the lack of small object feature information and improve the detection rate of small object. The experimental results show that the proposed method has satisfied results for small object detection under the premise of ensuring the accuracy rate P and recall rate R of the comprehensive object detection.

中文翻译：

---

基于改进深度学习的小目标检测方法

摘要

本文提出了一种基于反向残差结构（IR-PSN）的并行SSD（单发多框检测器）融合网络，以解决提取特征信息不足和深度学习小目标检测效果不理想的问题。首先，在SSD网络中采用反向残差结构（IR）代替池化层。改进后的SSD网络称为IR-PSN的深层网络，用于提取图像的高级特征信息。其次，构造基于倒立残差结构的浅层网络，以提取图像的低层特征信息。最后，将浅层网络与深层网络融合在一起，避免缺少小物体特征信息，提高小物体的检测率。P和召回率R的综合对象检测。