Object Detection Using Deep Convolutional Neural Network

Object detection is a widely used computer vision task which is the basis of some complicated vision tasks.Currently,it's a hot research area that using deep convolutional neural network in object detection.From pipeline,most models can be summarized into two categories,one-stage and two stage.The one-stage model is more popular in industry since its balance between speed and m AP(Mean Average Precision).The two-stage model is more powerful in vision competitions because of its excellent m AP.This paper conducts research with the framework of the classic one-stage model Single Shot Multi Box Detector(SSD).For general object detection,an in-depth study is carried out on how to increase m AP while keeping the advantage in speed.The contents are as follows:1.Proposed a feature fusion method(The Reversed Fusion Block with Attention,A-RFB).This approach uses fusion blocks to process pairs of feature maps from different layers.In one pair,the resolution of the feature map from a lower layer is exactly twice of the one from a higher layer.Consequently,it builds a reverse feature pyramid at the back end of networks by using several blocks.In each feature fusion block,the higher feature map is firstly expanded by deconvolution operation,and then fused with the lower feature map by Hadamard product.Finally,the fused feature will be recalibrated by an attention network.Through integrating the abstract semantic information from a higher layer feature map into a lower one,A-RFB improves the detection performance and significantly,increases the detection m AP on small objects.This paper avoids some possible redundant operations through structural principle analysis and experimental verification.Therefore,the structure of A-RFB is simple and effective.It can improve the detection m AP with only reducing 5% FPS by combining A-RFB with SSD.It's much faster than DSSD(Deconvolutional Single Shot Detector)which was proposed by the author SSD,and only 0.1% m AP behind.Moreover,the experiments in this paper prove that A-RFB is also able to improve the performance of lightweight one-stage models which use Dense Net and Mobile Net v2 as their backbone network.2.Inspired by the mixup method in the image classification field,this paper proposes a mix data augmentation for training object detection model based on SSD framework.For vicinal risk minimization,this paper sets a hypothesis that linear interpolation of images can be expressed as the linear interpolation of label vectors through feature transformation.By mixing two random training images,this method help model learn the vicinal relationships among the samples from different categories and then,improve the model's generalization capability.Through analyzing the experimental results and the principle of the model,this paper proposes a series of training rules such as weighting the class labels with the mixed coefficient,retaining the ground truth bounding boxes,setting a mix threshold,and combining with random crop.Using mix data augmentation method can improve m AP without any changing on network structure.In addition,training the A-RFB SSD with mix data augmentation will greatly increase m AP,which has surpassed serval improved SSD variants.