Research On Improved Object Detection Method Based On Convolutional Neural Network

Object detection is a basic but very important task of computer vision.Nowadays,the object detector based on deep learning algorithm,especially convolutional neural network,has become the mainstream of detection field.Compared with the detector based on the traditional feature extraction algorithm,it has a significant improvement both in the detection accuracy and reasoning speed.In the convolution neural network based object detector,the two-stage detector Faster R-CNN and the single-stage detector YOLOv2 and YOLOv3 have obvious advantages in detection accuracy or reasoning speed compared with other detectors in performance.However,we can still optimize and improve it to get more efficient detection results.There are three innovations in this paper:Firstly,in view of the inaccuracy of region proposals representing the potential positions of objects in images generated by the region proposal network of Faster R-CNN,which restricts the further improvement of the final detection accuracy,pure iterative refinement model and LSTM iterative refinement model based on Faster R-CNN are proposed.Pure iterative refinement model is trained on PASCAL VOC 07 trainval set.On test set,the best m AP obtained by the iterative model is 1.1% higher than baseline based on VGG-16 backbone network,and is 1.5% higher based on ZFNet.LSTM iterative refinement model can also improve the detection accuracy well.Moreover,the model can be back-propagated from the later region proposal iterative process to the earlier iteration.Thereby the end-to-end learning is realized.Secondly,aiming at solving the problem of sample imbalance in the training process of single-stage object detector based on deep learning,starting from optimizing the classification loss function of model training process,a new dynamic modulation cross entropy loss function based on Softmax classification is designed.This loss function can effectively reduce the loss weight of easily classified negative samples in the training process,and correspondingly improve the loss weight of difficult samples,so that the whole training process of the model becomes efficient.And from the loss function expression analysis to carry on the enough mathematics theory support.The dynamic modulation cross entropy loss function based on Softmax classification is designed to replace the standard cross entropy loss function in the loss calculation of classification prediction in the training of YOLOv2.The experimental results show that the application of this loss function can significantly improve the detection accuracy of YOLOv2.Finally,aiming at the problem of how to train the object detection model based on deep learning more efficiently,this paper explores the improvement of detection performance by using four kinds of optimization training methods.Data augmentation,class label smoothing,learning rate optimization and random scale training are applied to the training of one-stage detection model YOLOv3 and two-stage detection model Faster R-CNN on Pascal VOC dataset,which proves that all these methods can improve the detection accuracy to a certain extent.And explain the reasons why various methods can effectively improve the accuracy.The experiment of MS COCO dataset proves the generalization ability of these methods.Finally,it is shown that the object detection model can be trained more efficiently by using these optimization training methods.