Research On Object Detection In The Wild Based On Deep Convolutional Neural Network

Object detection is an important and fundamental task in computer vision and artificial intelligence,which aims to localize the position of each instance and classify the category of each target object.Moreover,it is a core research topic in the community of academia and industry.Although after decades of development,the current object detection algorithms only perform well in laboratory post images,but it usually relies large-scale dataset with annotations in real-world scenarios,and the detection results are far from satisfactory.Meanwhile,for the problem of small object(less than 32 * 32 pixels)detection in real-world scenarios,it is still in the stage of designing specific problem algorithm for each problem,and the robustness is poor.Towards those problems,in this paper,we conduct the research on the object detection under weakly annotated labels and missing labels,the feature learning method for detecting small objects and the small object detection framework.Our research can promote the object detection from the laboratory to the practical application.The main research contents of this paper are summarized as follows:(1)In view of the problem of existing object detection algorithms based on DCNNs usually rely on large-scale datasets to train a detection network,however,building such a large-scale dataset is a huge and time-consuming project,we propose an weakly-supervised object detection algorithm based on pseudo ground truth mining algorithm.In order to solve the problem of inaccurate positioning and low detection accuracy of the weakly-supervised object detector,we propose an weakly-supervised to fully-supervised object detection framework,in which the pseudo ground-truth excavation(PGE)algorithm can generate an accurate and tight pseudo ground-truth bounding box for each object.At the same time,a pseudo ground-truth adaptation(PGA)algorithm is proposed to fine-tune the pseudo ground-truth bounding boxes generated by PGE.Finally,an object detection network is trained by using the mined pseudo ground-truthes.The experimental results show that our proposed algorithm in this paper can overcome the problem of scarce training data in practical applications,and can greatly improve the accuracy of object detection compared with the state-of-the-art algorithms.(2)As for the problem of weakly-supervised object detection algorithms is worse than that of fully-supervised object detection algorithms,we propose an missing labels object detection algorithm based on incremental learning method.Firstly,we investigate the influence of missing labels in training images on general object detection algorithms,and then we give suggestion on how to construct a large-scale object detection dataset with limited human and material resources.Secondly,we try to improve object detection performance by introducing a small number of human annotated training data.Meanwhile,we treat this task as a missing label detection problem.We use the incremental learning framework to find the missed labels for each of the training sample from rough to accurate.Then,the mined labels and the partial annotated labels are combined to train an object detection network.Experimental results show that the proposed can overcome the problem of missing label in the object detection task.Our proposed algorithm can achieve high detection performance by using a small number of annotated labels,and we can reduce the gap between full-supervised and weakly-supervised algorithms.(3)When it comes to the difficulty of extracting features of small objects in the current object detection algorithm,we proposed a new multi-channel high-resolution feature extraction method.The parallel convolution layer in this algorithm can extract the deep high-resolution features of small objects.The learned features can overcome low semantic information in the shallow convolutional layers in the existing small object detection algorithms.In order to further finish the multi-scale object detection task,we propose a hierarchy feature fusion method,where high-resolution features owned rich detailed information are combined with low-resolution features have strong semantic information to build high-quality representations.Finally,experimental results show that the feature extraction method proposed in this paper can better adapt to the multi-scale characteristics,and can improve the performance of small objects detection in the real-world scenarios.(4)To solve the problem of existing object detection algorithms cannot detect small objects under their own limited details,and these algorithms also cannot reduce the influence of occlusion,illumination,and blurry.In this paper,we propose a novel end-to-end small object detection framework based on generative adversarial network.To our best knowledge,we are the first to apply the super-resolution technology in the task of small object detection.Specifically,GAN was used to generate the high-resolution images corresponding to the low-resolution images,and then the object detection task was realized on the clear generate high-resolution images.Furthermore,several new loss functions are designed to enable the generator network to produce clearer super-resolution images.Extensive experimental results show that the proposed algorithm can overcome the small object detection problem with insufficient details,and it can solve the problem of small object detection problem and reduce the influence of illumination,occlusion,blurry in the task of small object detection.