Research On High-quality Image Segmentation Based On Deep Learning

Image segmentation is a hot topic in computer vision research.With the rapid development of computing power of the hardware,image segmentation algorithms based on deep learning are widely used in digital video special effects and artificial advertising design to extract objects of interest.To enhance the visual effects in artificial design,high-quality image segmentation algorithms are necessary.The image segmentation algorithm predicts a discrete semantic label for each pixel so that it can extract objects with hard boundaries and compose them on new background images.But this cannot meet all the demands.For example,the composition results usually look unnatural with the hard segmentation results,especially for the images with motion blur,hair details,and transparent objects.Therefore,a high-quality soft segmentation method is needed to deal with this problem.The image matting algorithms use the trimaps as prior information to predict an alpha value in a float type for each pixel,which can be considered as soft segmentation results.However,these methods are less automated.Designing an automated and effective image matting algorithm is essential.In this thesis,we firstly propose the erroneous pixel prediction method for image segmentation which can predict errors and correct errors for the given segmentation network.This method consists of two stages: 1)Predict pixel-wise error probability for the given initial result.2)Re-estimate new labels for the pixels with high error probabilities.3)Fuse the initial result and the re-estimated result with respect to the error probabilities.The erroneous pixel prediction is formulated as a classification task and we employ an error-prediction branch in our network to predict the pixel-wise error probabilities.We also introduce another network branch called detail branch.This branch is designed such that the training process is focused on the erroneous pixels.We experimentally validate our method on public datasets and the experiment results show that our method can correct errors for various existing segmentation networks.Especially,our network can achieve 82.88% of mIoU in the Cityscapes testing dataset,which is 0.74% higher than DeepLabV3+.The automated image matting algorithm can benefit from the semantic features and coarse segmentation results generated by segmentation networks.The late fusion trimapless image matting algorithm in this paper takes a single RGB image as the input and outputs the alpha matte.This method firstly uses two segmentation branches to predict the foreground and background segmentation respectively.Then a fusion branch is exploited to integrate the two branches' results.In particular,through the design of loss function,the prediction results of the two segmentation branches can be considered as the upper bound and the lower bound of the true alpha values respectively.We validate the effectiveness of our two-branch design through experiments and compare our method with other existing methods on public datasets.Our late fusion method outperforms the state-of-the-art methods on the human image matting dataset.