Research On Object Detection In Drone-view Images Based On Deep Convolutional Neural Networks

Object detection on drone-view images has become a research hotspot in computer vision in recent years,owing to the fact that drone technology has been widely applicated in agriculture,express delivery,intelligent city and other fields.Compared with natural scene images,drone-view images have the following features:high resolution,complex background,variety of camera angles,large scale changes and so on.Traditional detection algorithms based on handcraft features can not meet our accuracy requirements.Deep learning technology has become the mainstream research direction.This paper focuses on the detection of small object,complex background and global information modeling in drone-view images and the main work and innovations of this papare are summarized as follows:(1)A new feature fusion network named Inverted and Decoded FPN,which is based on FPN,is proposed in this paper to solve the problem that there are a large number of small objects in drone-view images dataset.ID-FPN can effectively fuse feature maps at different levels from convolutional neural networks and enhance semantic information.Experiments show that ID-FPN can be used as an independent feature extractor in existing detection algorithms.ID-FPN helps Faster R-CNN to improve the ability of small objects detection on VisDrone dataset,which is about 1.5%higher than FPN in terms of mAP.(2)Non-local Block is introduced to model the global context information of drone-view images.In general,global context information is great helpful to detection task.But existing detectors detect each object independently,which makes the relationship between objects and global context information not considered.Non-local Block enables the detectors to capture the context information of targets and effectively reduce false alarm rate and false positive rate.(3)Aiming at solving the problem of complex background,this paper proposes an object detection algorithm:Attention Based Faster R-CNN.The new algorithm designs a new attention module on the basis of Receptive Field Block,and introduces ID-FPN and Non-local Block module.Attention mechanism makes convolutional neural networks be able to remove redundant background information and enhance the ability of feature representation.The experimental results show that new object detection framework can achieves leading performance on VisDrone dataset,which is 1.2%higher than other two-stage detection algorithms.