Research On Object Detection And Instance Segmentation Method Based On Unmanned Aerial Vehicle Images

Recent advances in the Unmanned Aerial Vehicles(UAVs)and the wireless image transferred technology,have consistently promoted the development in aerial photography of UAVs,as the eye of the earth,to facility urge demand in many domains.UAV patrols not only play a key role to reduce the consumption of labor and resources,but also guarantee the safety of operations in some certain traditional industries.Therefore,how to locate the target accurately and analyze the spatial location relation is the key task of most UAV patrol operations.Traditional object detection and segmentation methods adopt a two-stage framework to improve the accuracy.In general,a backbone network is firstly utilized for feature extraction and region proposal generation.Then,a multi-task network branch is leveraged.This kind of two-stage object detection method may fail to detect small objects.Details of an input image will be lost especially in dealing with high-resolution images.It thus more suitable for low-resolution image processing which is collected in natural scene with conventional perspective,but fails to perform small object detection in high-resolution UAV images.In order to the above-mentioned problem,in this thesis,a clustering-based pre-processing algorithm and a multi-task model for both object detection and instance segmentation are proposed to improve the performance.Finally,a coarse-to-fine searching and detecting strategy is presented for small object detection in UAV images.Specifically,the core idea of the proposed method can be summarized as follows:(1)The pre-processing algorithm aims at searching potential areas with objects from high-resolution UAV images.Areas with small objects will be zoomed in based on the result produced by clustering algorithm.(2)In order to preserve the low-level information for small objects in high-resolution images,a multi-scale parallel subnetwork structure in incorporated for restoring the location information for small objects under a high-to-low resolution framework.(3)To overcome the imbalance of positive and negative samples for small object detection,a coarse-to-fine region proposal network is adopted to gradually improve the quality of positive samples in region proposal generation.(4)A double scoring mechanism is proposed for both considering detection and segmentation branches.According to the predicted bounding box and mask,bounding box Io U(Intersection of Union)and mask Io U score can be calculated based on ground-truths.Region-level and pixel-level classification accuracy can be further improved by leveraging the proposed double scoring strategy.To demonstrate the feasibility and efficiency of our proposed method,in thisthesis,a large UAV dataset for object detection and instance segmentation is constructed.Pre-processing algorithm and multi-task model for small object detection and instance segmentation are implemented.Experimental results show that our proposed coarse-to-fine small object searching and detection method is superior to the traditional two-stage detectors.At the meanwhile,more accurate detection and segmentation results can be used to calculate the distance and spatial localization more accurately,and the results also show the availability and effectiveness of the proposed method.