Small Weak Object Detection In High-Resolution Remote Sensing Images Based On Deep Neural Network

In recent years,detecting small weak objects in optical high-resolution remote sensing(HRRS)images has become a hot issue in remote sensing image processing.The small weak objects usually take up small proportions of an HRRS image and are with their sufficient variations in color,shape,and texture so that the objects' features are easily affected by weather,illumination,and occlusion.These characteristics of small weak objects make their detection a more challenging task than generic object detection.In this paper,a series of works have been carried out to address the three problems of small scale,weak feature response,and limited annotation samples in small weak object detection.These research works can meet the application needs of urban monitoring and disaster prevention,which also have significant scientific and social significance.The main innovations and research contents of the paper are summarized as follows:(1)This paper first introduces the research background and the challenges of small weak object detection in HRRS images in detail.It then systematically reviews representative benchmark datasets,mainstream detection methods for HRRS object detection,and the current research statuses for addressing different challenges in small weak object detection.The above datasets and detection methods are analyzed to summarize their contributions to small weak object detection in HRRS images.Furthermore,to solve the problem of lack of challenging unmanned aerial vehicle(UAV)object detection dataset in the remote sensing community,this paper proposes a 10-category unmanned aerial vehicle object detection dataset(UAVOD-10),which reflects the coexistence of multiple-type objects,large differences of object appearances,unbalanced distribution,and noise in the real-world.The proposed UAVOD-10 data provides a data basis for the related research works.(2)The paper proposes a sampling-balance-based multi-scale cascade network(SB-MSN).Due to the small scale of small weak objects and the considerable differences in the ratio of the objects of interest to the background,there are many lowquality negative instances in the training process of deep-learning-based detectors,which causes poor detection performance.The proposed detection method uses a multiscale information module to extract objects' visual features,then develops an intersection-over-union(Io U)based sampling balance strategy to filter low-quality negative examples,finally applies a multi-stage detection head to optimize the predictions progressively.As a result,SB-MSN can efficiently detect small-scale ground objects.(3)This paper designs a context-scale-aware weak target detector(CASDet)to detect objects with high intra-class differences and noise.CASDet tries to mine the global and local context information of the objects of interest to enhance their feature response strength.Then,an augmented pyramid network is combined to extract the objects' multi-scale visual features and fuses the cross-scale information from the lowlevel to the high-level features.The proposed detector effectively improves the detection accuracy of weak feature objects.(4)To solve the lack of annotation samples in small weak object detection,a deeplearning-based semi-supervised sample generation framework(SSGF)is proposed.SSGF combines the deep learning features,a self-label technique,and a discriminative evaluator to complete the task of semi-supervised learning.SSFG can transform the unlabeled samples to an annotation set.Therefore,the problem of limited annotation samples is addressed.Focusing on detecting small weak objects in HRRS images,this paper first proposes a challenging UAV dataset,UAVOD-10,then develops a series of methods to solve the problems of small scale,weak feature response,and limited annotation samples in small weak object detection.The proposed approaches and data are critical to promoting the development of small weak objects in HRRS images.