A Lightweight Object Detection Method in Aerial Images Based on Dense Feature Fusion Path Aggregation Network

Liming Zhou,Yahui Li,Yinghao Lin,Xiaohan Rao,Xianyu Zuo,Baojun Qiao

doi:10.3390/ijgi11030189

Liming Zhou, Yahui Li + Show 4 more

Open Access

https://doi.org/10.3390/ijgi11030189

Copy DOI

Abstract

In recent years, significant progress has been obtained in object detection using Convolutional Neural Networks (CNNs). However, owing to the particularity of Remote Sensing Images (RSIs), common object detection methods are not well suited for RSIs. Aiming at the difficulties in RSIs, this paper proposes an object detection method based on the Dense Feature Fusion Path Aggregation Network (DFF-PANet). Firstly, for better improving the detection performance of small and medium-sized instances, we propose Feature Reuse Module (FRM), which can integrate semantic and location information contained in feature maps; this module can reuse feature maps in the backbone to enhance the detection capability of small and medium-sized instances. After that, we design the DFF-PANet, which can help feature information extracted from the backbone to be fused more efficiently, and thus cope with the problem of external interference factors. We performed experiments on the Dataset of Object deTection in Aerial images (DOTA) dataset and the HRSC2016 dataset; the accuracy reached 71.5% mAP, which exceeds most object detectors of one-stage and two-stages at present. Meanwhile, the size of our model is only 9.2 M, which satisfies the requirement of being lightweight. The experimental results demonstrate that our method not only has better detection accuracy but also maintains high efficiency in RSIs.

Highlights

With the progress of Remote Sensing Images (RSIs) sensors, people can obtain high-quality and high-resolution aerial images by using remote sensing technology
We conducted some experiments on the DOTA dataset to verify the validity of our proposed method and compared our method with current popular one-stage object detection methods, such as SSD [22], YOLOv2 [11], RetinaNet [25], Adaptive Feature Aggregation
Two-stage object detection methods include RetinaNet [25], Rotation-sensitive Regression Detector (RRD) [47], Rotation Sensitive Detector (RSDet) [48], Dynamic Anchor Learning (DAL) [49], Refi Ned Single-Stage Detector (R3Det) [50] and RepVGG-YOLO [51]. It can be seen from the table that the method we proposed achieves the best result among all the comparison methods, with mAP reaching 93.3%, which is 1.8 percent higher than the suboptimal method (RepVGG-YOLO)

Summary

Introduction

With the progress of RSIs sensors, people can obtain high-quality and high-resolution aerial images by using remote sensing technology. Target detection in RSIs is of great significance in military, civil and other aspects. Deep learning has promoted great progress in various computer vision problems, for instance, object classification [1–3], object detection [4–6], object tracking [7,8]. The application of deep learning models to aerial object detection has aroused more and more attention. For the past few years, CNNs has emerged in many object detection algorithms, which have obtained good results both in speed and accuracy. Compared with traditional object detection methods, for example, Deformable Parts Model (DPM), Histogram of Oriented

Methods

Results

Discussion

Conclusion