Weakly Supervised Vehicle Detection In Remote Sensing Images Based On Multi-instance Learning

Vehicle detection and tracking in remote sensing images play an important role in civil and military applications,such as highway traffic monitoring,management,and urban transportation planning.Compared with the target detection in the ground view image,vehicle detection in the remote sensing image is a good supplement for vehicle detection,which can obtain a more global perspective of urban traffic conditions.However,due to the simple appearance of the remote sensing image,complex background and small size,etc,we are faced with many new problems,such as the lack of apparent and detailed features,the background interference,and the difficulty in labeling.In view of the difficulty of annotating a sufficient number of satellite image instances across different resolutions and imaging conditions,weakly supervised vehicle detection become very important for satellite image analysis and processing.To cope with this challenge,we present a weakly supervised scheme for the training of smart and labor-intensive vehicle detectors in large-scale application.Our scheme only requires a region-level group labeling.That is to say,whether this area contains vehicles without explicitly denoting the vehicle’s rectangular boxes.To this end,we designed a novel weakly supervised,multi-instance ranking algorithm.Towards progressively learning instance-level classifiers based on the efficiency issues of weakly supervised tags in vehicle detection,we introduce extra low-cost tagging information as a prior,and then use the multi-instance learning algorithm to classify and locate vehicle instances under this couting prior.We compared with some common vehicle detection algorithms and achieved good results.In summary,the contents of our research mainly include the following three aspects:(1)We use a user-friendly,lightweight and regional-level annotation to collect large-scale vehicles tagged with a variety of sensor and image conditions,and use multi-example progressive learning for vehicle detection based on such weak annotations..(2)Due to the inaccuracy of weakly supervised positioning,we have introduced the context information of remote sensing into the vehicle detection of ground-level perspectives,which are used for the detection of remote sensing small targets.The model learned through the migration of this information can be used to effectively for detection.(3)By using multi-instance learning to progressively learn vehicle detectors from weak labels,we conduct a useful exploration of the remote sensing image vehicle detection in a portable way,with better trade-offs and improvements in accuracy and recall rates.