Research On Image Semantic Segmentation With Transfer Learning

In recent years,significant progress has been made in the field of supervised image semantic segmentation based on convolutional neural networks.To solve the labor-intensive and time-consuming issue of manual labeling,one of the common solutions is to collect similar images through gaming videos for automatically-generating labels,and then apply transfer learning for migrating the model trained in the synthetic scene to the real scene.Due to the domain shift,simply applying the model learned on the synthetic scene(source domain)to the real scene(target domain)will incur considerable performance drop.This thesis focus on the application of transfer learning for image semantic segmentation and the main contributions can be stated as follows.(1)We propose a novel semantic segmentation method based on entropy minimization.By introducing theentropy minimization in the tareget loss,the segmentation network can be effectively trained over the target domain data set,and the generalization ability can be improved.However,considering that the training process is often unstable,the method based on entropy minimization may cause the model predictions to be biased toward a single category in the data set.Thus,we propose to use the twin network structure for guiding the trained segmentation network toward converging.Experiments show that the proposed method can significantly improve the final accuracy for various semantic segmentation tasks.(2)We propose to use the atrous convolution for designing domain discriminator.The performance of the discriminator network has a great influence on the overall performance of the semantic segmentation.Therefore,we use the atrous convolution to replace the traditional convolution-based domain discriminator,which expands the receptive field without additional training parameters.The experimental results show that the proposed domain discriminator can significantly improve the system performance for transfer-learing-based semantic segmentation transfer.(3)We introduce image style conversion technology for source image preprocessing.The feature domain adaptation method partially solves the problem of domain shift.However,the difference in vision(such as shape,light,etc.)between the synthesized picture and the real picture makes the model very difficult to train.To this end,we use the attention space pyramid pooling structure to design an image style conversion network,which reduces the pixel-level difference between source pictures and target pictures.Experiments demonstrate the effectiveness of the proposed method.