Research On Multi-Scene Key Technologies On Image Transfer Learning

Convolutional neural networks are widely used in various image processing tasks,which has achieved good results in many tasks such as image classification,objects detection and semantic segmentation.In many applications,the convolutional neural networks need a large amount of annotated data for supervised learning.It is expensive to annotate these data and it requires that the data doesn't change after a period of time has passed.In other words,it requires that the training set and test set are independent and identical distributions.However,as the environment changes,the training data set needs to be relabeled in many applications.Therefore,many researchers try to solve this problem through transfer learning.In this paper,we did some researches on transfer learning based on convolutional neural networks.In detail,the above contributions are as follows:(1)color consistency correction of remote sensing image based on color transferringFor the color correction problem of remote sensing images,there are big resolution gap between the image to be corrected and reference image which are from high-resolution satellites and low-resolution satellites,respectively.It is difficult to obtain an accurate pixel-level correspondence between the image to be corrected and the reference image,which makes it impossible to obtain training data that can be used for strongly supervised learning.In order to solve this problem,a deep convolutional neural network model based on weakly supervised learning is proposed in this paper.Firstly,a variational autoencoding network model is trained by unsupervised learning,which can extract hierarchical features of images.Then,some style parameters are trained to control the color of image generated by autoencoder network model.Finally,the color of image is corrected by autoencoder network and style parameters.The experimental results show that the proposed method is better than UNIT and Cycle GAN methods on the indexes the paired points of SIFT feature,gradient direction loss and EM distance.(2)night-time road scene parsing based on color transferringWith the absence of night-time annotated data,it is difficult to parse the night-time road scene for the convolutional neural network trained on day-time annotated data.In order to solve this problem,we improve the accuracy of model by color transferring,which can reduce the color difference between day-time and night-time data.Firstly,we did some studies on the common convolutional neural networks about scene parsing,and proposed an adaptive upsampling method to optimize the objects' edges of semantic segmentation results,which made a better results.Then,we did some studies on color transferring about night-time road scene parsing,which aims to improve the accuracy of model by reducing color differences between night-time images and day-time images.Finally,a scene parsing model is trained to parse night-time images using day-time data.The experimental results show that color transferring can improve the performance of model.Compared with the case without transfer learning,the accuracy is improved by 1.9% on m Io U.(3)night-time road scene parsing based on semantic transferringAlthough the color transferring can achieve better results on night-time road scene parsing,there is a problem that the joint optimization can't be performed on transferring model and scene parsing model,which may be an obstacle to obtain further improvement.To this end,we try to parse night-time road scene by semantic transferring,which can merge transfer model and scene parsing model and perform joint optimization in one phase.Firstly,a feature extractor that is not sensitive to lighting is trained by unsupervised learning,which can reduce the difference between day-time and night-time data in the semantic space.Then,a scene parsing model is trained on day-time annotated data feed with feature data in semantic space.Finally,the model which composed of transfer model and scene parsing model is used to parse the night-time road scene.The experimental results show that there is a improvement of 5.24% on the m Io U index after using the proposed semantic transferring model.In summary,for the color consistency correction of remote sensing images and the night-time road scene parsing,we did some studies on transfer learning.Many quantitative and qualitative experimental results show that the proposed methods can achieve better results on the above two problems,which indicates our researches are valuable in some potential applications.