Research On Partial Configuration Model Based Partially Occluded Object Detection Methods In High-resolution Remote Sensing Images

Object detection is one of important tasks and research hotspots in remote sensing image interpretation.With the great development of sensors,the spatial resolution of remote sensing images is getting higher and higher,and the demand for refined object interpretation makes the frequently happened problem of object occlusion in remote sensing images increasingly prominent.How to quickly find as many suspicious objects as possible from complex backgrounds and accurately infer the true location of full object from unoccluded evidence are the difficult points for partially occluded object detection(POOD).However,state-of-the-art object detection methods regard occlusion as specific kind of unstructured image noises,making it unable to adapt to the common severe occlusion in high-resolution remote sensing images(HR-RSIs)and obtain accurate fullobject locations under occlusion.To solve the above problems,this dissertation uses the deformable part-based model(DPM)as a theoretical basis,and innovatively proposes a partial configuration model(PCM)specially oriented for POOD.Through modeling the local features of objects,the spatial and evidential interrelationships between local partial configurations and the full-extent object are built.The fusion of multiple partial configuration evidences will then contribute to the description of the object occlusion state,which ultimately achieves high detection recall,high locating accuracy and high computational efficiency under partial occlusion.The proposed PCM method in this dissertation obviously improves the POOD performance of existing detection algorithms,including the deep learning based methods.The main work and the innovations of the dissertation are listed as follows:1.A partial configuration model(PCM)for occluded object detection in HR-RSIs is proposed.We analyze the inherent causes of the unsatisfactory performance of the existing object detection algorithms in details when facing occlusion,and innovatively propose a two-level object model,which is called partial configuration model.This model uses an additional partial configuration layer to block the effect of occlusion from passing on to the object detection layer,so that the partial configuration model would have sufficient evidence to confirm the existence of partial occluded objects.At the same time,PCM can accurately infer the actual location of the occluded object from the hypotheses from the partial configuration layer.In this dissertation,the model is verified on multiple datasets and multiple object categories.Experimental results show that the proposed PCM model obtains very good detection performance under occlusion and achieves great performance improvement compared to conventional object detection algorithms.Based on the intermediate results of PCM,we can also infer the occlusion states of all detected objects.2.An automatic and fast PCM model generation method(AFI-PCM)based on part sharing mechanism is proposed.However,PCM needs to manually specify semantic parts and skeleton graph for each category of object at the initial stage of training.In order to tackle this,this dissertation proposes to share the parts between the partial configuration models.In this mechanism,a shared set of parts from a trained full-object DPM model are used to automatically design category skeleton graph and each partial configuration,and new models of partial configurations are finally assembled and obtained.This greatly reduces sample labeling and model training volume,and makes the fully automatic and fast training of PCM become possible.The experimental results demonstrate that the training speedup of the AFI-PCM method is 6.7x and 2x,respectively,for the object category of aircraft and ship,compared to PCM.Moreover,the more elements designed in the partial configuration layer in the original PCM,the greater the acceleration ratio we will obtained.3.A unified partial configuration model framework(UniPCM)for fast POOD is proposed.The sharing of parts during the training phase only speeds up the design and training of the model,which also requires an additional model assembling process.In the more important stage of detection,the part sharing mechanism is actually not used.This dissertation proposes a unified partial configuration model framework UniPCM,which incorporates the entire model training and detection into a unified framework.During the training phase,the shared part filters are directly used for intra-PCM model weights learning and inter-PCM balance.In the detection stage,all POOD work is performed on the basis of the same set of shared part filter responses,which greatly reduces the amount of model calculation.At the same time,we also propose a method to directly estimate the spatial interrelationship from the deformation information of the trained DPM model,as well as a unified model parameter learning method.The training speed of UniPCM has been further accelerated with respect to the innovation point 2,and the model detection speed obtaines a speedup over 10 x,while at the same time the performance of UniPCM is comparable to the original PCM method.Compared to the deep learning-based object detection method MS-CNN,the proposed algorithm has a great advantage in detection accuracy on both occluded airplane and ship datasets.4.An accurate non-maximum suppression(NMS)method is proposed for object detection in HR-RSIs.In HR-RSIs,there are usually a large number of clustered objects,such as ships and vehicles.There is thus often a large overlap rate between their bounding boxes,which leads to a situation that conventional NMS methods cannot accurately find those relatively low-scored real thus suppressed object,and the object locating accuracy would decrease.Some NMS methods try to obtain higher locating accuracy through ingroup hypotheses regression,which also suffer from the problem of hypotheses number imbalance.To solve above problems,this dissertation proposes an accurate NMS method.It changes the traditional strategy of directly selecting the maximally scored object and deleting other objects,and adopts a gradually weight decreasing strategy for the non-maximally scored objects to ensure that these sub-maximum but real objects can also be found in the subsequent iterations.At the same time,when performing the in-group clustering,weighted fusion and score magnification techniques are used to obtain more accurate object locating results.This method is simple and easy to implement and does not require model training at all.The experimental results on the dataset containing ten categories of object show that this method achieves obvious performance improvement in most categories.