Research On Incremental Learning Of Object Detection Algorithm Based On Convolutional Neural Network

Object detection in images is one of the core tasks in the computer vision field and the cornerstone of computer understanding images.It can be applied to intelligent video surveillance,automatic drive,robots,smart home,medical image assistant diagnosis and other needs,products.Since the birth of AlexNet,the champion of the ImageNet Image Classification Competition in 2012,the Convolutional Neural Network(CNN)has been widely used in the computer vision field.Deep learning technologies especially CNN also greatly improve the accuracy of object detection task.With the advantages of CNN and the unremitting efforts of a large number of scholars,nowadays computer programs can almost match human vision in object detection tasks.Some representative works were produced like R-CNN,Fast R-CNN,Faster R-CNN,YOLO,FPN and RetinaNet.In specific scenarios,the accuracy of object detection algorithms is gradually improved to the saturation nearly.But some new problems are gradually coming into people's vision.Incremental learning in computer vision is one of these emerging problems.Object detection is the main task of computer vision.Therefore,this thesis is to research the incremental learning of obj ect detection.At present,the most accurate methods in the object detection field are all based on CNN.However,the connection type methods such as neural networks have catastrophic forgetting problem.This problem is:if a new task is trained on an existing neural network model,the existing ability of this model will be lost sharply.Therefore,the CNN-based object detection algorithms also have such problem,that is,when adding new detectable categories to an existing object detection model,if the existing model is directly trained with the data of new categories,the final model will lose the detection ability on the old categories to a great extent.The goal of this thesis is:how to better add new detectable categories to an existing object detection model.Two methods are researched in this thesis.The first method is based on the idea of feature extraction.This is an intuitive way to add a proprietary layer to new categories.The contribution of this thesis is:proposed the details which can improve accuracy when using this method,and researched some properties of this method through experiments.This thesis puts forward:(1)Moving forward the adding position of the new categories' proprietary layer can significantly improve the detection accuracy on the new categories.(2)Adding shallow bypass layers for the new categories' detection can improve the detection accuracy on the new categories.The second method is retaining loss function method,and it can be regarded as proposed by this thesis.It is based on LwF(Learning Without Forgetting).LwF is an incremental learning method for image classification proposed by Zhizhong Li et al.LwF drew inspiration from model compression method,which is widely known as model distillation.The contribution of this thesis is to take the lead in applying LwF idea to popular object detection algorithms which have Regional Proposal Network(RPN)module.And this thesis got a higher accuracy baseline on this task than state of the art methods.Through a large number of experiments,it is worth considering various details when adding new detectable categories to the object detection model with the idea of LwF.In this thesis,many details are experimented and the best ways to deal with these details are selected.These details will have a certain impact on the final accuracy of the model.