| Object detection,which locates the location of the object in the video sequence or image and determines the category of the object,is a research hotspot in the field of computer vision.In intelligent video surveillance,driverless technology,computer-aided diagnosis technology,face Identification,human flow monitoring and other aspects,object detection has important applications.However,due to the influence of many factors such as deformation,occlusion,observation angle and environmental changes in the actual scene,object detection is still a very challenging task.How to design features that can accurately identify the target object without being affected by various external diversity factors has become the focus of research in this field.This paper innovatively proposes an object detection algorithm based on hyper feature pyramid and adversar learning: A-HFPN.In the feature extraction stage,the concept of hyper feature pyramid is proposed.The outstanding advantage is that it can fully extract object features from dfferent convolution layers and different sizes.In addition,an improved RPN network is also proposed,which is reasonable.The strategy predicts small objects at low levels and predicts large objects at high levels,making the entire RPN network better.At the same time,according to the susceptibility of object detection,a Mask network is proposed,which actively generates data by adding occlusion networks.It focus on a restricted space for generation: occlusion,instead of generating data in the entire pixel space directly.This paper also proposes a fine-tuning network,which uses the combination of RoI Pooling and RoI Align to effectively improve the detection accuracy.Finally,the final boudind boxes is filtered by soft-NMS.Under the Resnet-101 network architecture,the algorithm achieves the mean average precision of 81.1034% on the Pascal VOC 2007 test set and 73.52% on the DETRAC data test set,reaching the current state of the art detection level. |
PDF Full Text Request