Research On Transfer Learning Technology Based On Deep Convolutional Networks

Transfer Learning is a machine learning method that uses existing knowledge to solve different but related domain problems.The goal is to transfer existing knowledge and experience from the source task to the target domain.Deep convolutional network(DCN)is inspired by the mechanism of bio-visual cognition,it contains multiple-layer convolution and pooling structures and has been successfully applied to image recognition,object detection and other fields.This thesis studies and implements the application of transfer learning in image classification category expansion and object detection extension based on the deep convolutional network.The main work is as follows:1.For the effects of various data augmentation methods on image classification tasks of deep convolution networks,a variety of data augmentation methods are used to extend training image samples,and a subset of ImageNet and CIFAR-10 are selected as the original data sets,and AlexNet is used as the pre-training network model.The experimental results show that,under the same conditions,the test accuracy of small-scale datasets improve significantly after data augmentation,and the overall performance of the four single-item augmentation methods(Crop,Flip,WGAN,Rotation)is superior to other methods,and the appropriate combination is more effective than the single item.2.For the category expansion problem in image classification tasks,we have researched and implemented the image classification category expansion algorithm based on fine-tuning and feature extraction.The former combines the pre-training model and the layer freezing method to fine-tune the model to achieve the category expansion,the latter uses the pre-training model as the feature extractor to extract the feature of the corresponding layer and train the SVM classifier to achieve the category expansion.The experimental results show that when the number of newly added category samples is sufficient or equivalent,the fine-tuning method should be used to achieve category expansion.When the number of samples is small,the pre-training model is used as a feature extractor to train the SVM classifier to achieve category expansion.3.In order to solve the problem of target expansion in object detection,an object detection extension algorithm based on SSD model fine-tuning is studied and implemented.By fine-tuning the SSD-VOC model with the target data set,obtains the fine-tuned model,and the fine-tuned model is concatenated with the original SSD-VOC model to achieve target expansion.The experimental results show that the fine-tuned model can achieve 86.8%correct detection rate for the two types of targets in the target domain,and the concatenation model can achieve object detection from 20 to 22 categories.