Cascade Structure-based Deep Learning Object Detection Algorithm

Object detection is widely used in many area,it's an important part in computer vision.It is meaningful to analyse the accuracy and stability of detection algorithm.In this work,we mainly focus some chanlleging problem of object detection,we give a new idea for the design of object detection struture.This paper carries out the following work:Firstly,the paper introduces the development of detection alogorithm.The fundamental strcture of detection algorithm were studied in this paper,including fully connected layer,convolution layer,pooling layer,deformable operation,normalization method and RoI Align.we also introduce common module of detection algorithm,which include backbone network,anchor,loss function,ResNet-50 and ResNext-50 are base structure used in our paper.Secondly,R-CNN,Fast R-CNN and Faster R-CNN are introduced one by one.This paper reviews the current research results on multi-scale object detection,discuss some excellent structure in FPN,MatrixNet and TridentNet.The COCO dataset and Tensorflow training framework commonly used in the field of object detection are explained in detailThirdly,a novel data augment method is proposed,which can balance the distribution of dataset.Based on feature pyramid network,we proposed a bidirection feature pyramid network and attention pooling structure?The combination of Bi-FPN and attention pooling,called feature fusion network,can significally enhance the feature extraction process,improve the detection performance in multi-scale objects.Finally,based on detection result of feature fusion network,we proposed a cascade resample structure.The fact that detector IOU threshold,RPN input IOU and detector output IOU has a close relationship,When the detector IOU threshold matches the input IOU,the detector can get best result.By the cascade resample structure,we realize resampling the IOU distribution of the sample.Our model achieves 48.1%mmAp in COCO dataset when use ResNet-101 skeleton.Based on the algorithm proposed above,this paper uses TensorRT to accelerate the model,and uses Tensorflow serving and Qt to build an object detection application,which realizes the detection and visualization of local pictures without relying on the Python environment.