Research On Object Detection Methods Based On Deep Learning

Object detection is one of the most important research topics in the field of computer vision and the cornerstone of other advanced image understanding tasks.The task of object detection is to find the position of the targets of interest in the input images and recognizes the classes of the targets by using the object detection algorithm.In recent years,object detection algorithms based on deep learning have developed rapidly with the development of deep learning,especially convolutional neural networks.They are considerably strong as far as the feature extraction capabilities are concerned.Therefore,this paper conducts an in-depth research on object detection networks based on deep learning to ensure improve detection performance and detection practicability.Some issues that come along with constant innovation of object detection algorithm should be highlighted.Firstly,when the object's scale varies widely,the object detection algorithm is susceptible to scale variation.Secondly,the accuracy of object detection isn't high in the small samples,especially in complex scenes.Thirdly,the application of object detection on embedded devices remains to be improved.To this end,this paper conducts a series of studies on detection algorithms and applications in response to the above issues.This paper is summarized as following:(1)In order to address the problem of scale variation of detection algorithm,this paper proposes,based on the study of receptive fields,an object detection algorithm which enhances feature with a cross-depth convolution.Since the algorithm in this paper is based on the YOLOv3 algorithm whose feature pyramid module ignores the differences in the contribution made by different detection layers to object detection and leads to the problem of low object recall with large scale variation,this paper replaces it with newly-designed cross-depth feature enhancement modules.Each detection layer of the algorithm in this paper uses different feature enhancement modules for different receptive fields corresponding to them,which improves the adaptability of detection layers of each scale to the scale variation of the objects,as result,different detection layers make different contributions to object detection.Experimental results show that the algorithm in this paper can better solve the problem of object's scale variation.(2)It is found that the number of images in the training samples is the key reason that affects the accuracy of object detection algorithm.The network based on a small samples can easily learn the noise information and ignore the essential information in images,which leads to over-fitting.Aiming at this problem,this paper proposes an object detection network using class supervised method based on small samples,which consists of two branches,one of the branches is the basic detector,and the other is the class supervised network.The class supervised network is used to obtain the essential feature representation of per target class,and this feature representation is used for providing class supervised information for the basic detector.This class supervised process makes network pay more attention to the essential features related to per target class and ignore the secondary features of per target class.Experiments show that the algorithm achieves good results based on small samples.(3)Aiming at the embedded application of object detection algorithm,this paper designed a lightweight real-time detection network called Tiny SSD in the detection of UAV in a practical application project.The network improves the backbone network of SSD,and uses a cross-depth convolution feature enhancement module to implement the algorithm deployment in real-time embedded systems.Experimental results show that the algorithm can detect UAV very well on embedded devices,and the speed can be close to real-time.