Research On Bi-branch Object Detection Based On Mask Network And RPN

Object detection is a task with difficulty in detection in computer vision.It is close to the needs of real-life applications and plays a pivotal role in many applications such as human-computer interaction,robot vision,and car-assisted driving.In recent years,the research on computer vision in deep learning has made significant progress and has been applied to some areas of life.However,object detection has been widely studied as a branch of it.However,because of the variable scale of the object in the image,it faces enormous Research challenges.In the actual application scenario,the scale of the object varies greatly,and the objects of small,medium and large scales all have different numbers of distributions.However,most of the current object detection methods do not consider the influence of the scale range on the detection accuracy,resulting in these objects.Detection methods are often only adapted to object detection scenarios at certain scales.In this paper,the object detection problem is studied in detail,and a series of detection methods are designed to improve the detection of small-scale objects,the detection of mediumand large-scale objects,and finally design a method for multi-scale object detection.The research ideas of this paper are as follows:Firstly,under the inspiration of excavating occlusion samples,an occlusion network that can automatically generate occlusion samples is designed to improve the detection effect of medium and large scale objects.Next,a theoretical analysis of the reasons for the difficulty of detecting small-scale objects is carried out,and it is considered that the larger receptive field will ignore the amount of characteristic information that the small object originally has.Therefore,it is assumed that,under the premise of fixed filter size,small objects may be more suitable for shallow networks(less convolutional layers),thereby experimentally analyzing the influence of receptive fields on small-scale object detection,and looking for fixed filtering.Finally,according to the above research,a two-channel multi-scale object detection network structure based on occlusion network and regional suggestion network is designed.At the same time,the object size size and aspect ratio of Pascal VOC data set are counted,and the anchor pointspecification of the network is set.The research goal of this paper is to find the main factors affecting the detection effect of small,medium and large scales.Based on this,the current advanced detectors are improved to improve the detection results of different scale objects,and a deep method suitable for multi-scale object detection is comprehensively established.In order to verify the rationality and effectiveness of the analyzed factors and design methods,this paper carries out a large number of ablation experiments and comparative experiments on the COCO dataset.The results of the ablation experiment show that the design of the occlusion network greatly improves the detection of mesoscale objects and improves the detection of large-scale objects.At the same time,the ablation experiment also show that the frame small channel design is indeed conducive to the detection of small objects.The experimental results show that the proposed two-channel multi-scale object detection method and the advanced object detection method have strong competitiveness in scale,and excellent in the detection of small,medium and large scale objects.At the same time,the research theory of this paper is universal and can be applied to other object detection fields,such as multi-scale pedestrian detection and multi-scale vehicle detection.