Fast And Robust Algorithms On Semantic Image Segmentation

Semantic image segmentation or scene understanding is one of the key prob-lems of computer vision,which aims at labeling each pixel of an image given the known category set.In many applications,such as autonomous driving,portrai-ture,augmented reality and so on,an accurate and efficient semantic segmentation algorithm is indispensable.Thus,it becomes more and more interesting on how to implement fast and robust algorithms for semantic image segmentation.It is the main task of computer vision that paves the way towards complete scene understanding.In the most recent years,many researchers have proposed vari-ous approaches o tackle this critical problem,including traditional machine learn-ing methods and image feature technique,and recently widely used deep learning approaches.Based on various technique foundation and application scenarios,this paper introduce three kinds of approaches for semantic image segmentation.Specifically:Firstly,we introduce an image semantic segmentation algorithm based on la-bel transfer technique.The algorithm infers segmentation result of query image via transferring labels from annotated images that are nearest neighbors.Our work focuses on optimizing the search accuracy and speed of annotated images,and proposes an approximate nearest neighbor(ANN)search algorithm called Sparse Product Quantization.Combining the idea of soft assignment and the technique of product quantization,the introduced approach is able to not only reach a lower bound of quantization error and thus more accurate result,but also cost less com-putation of expanses.To show the efficacy and efficiency of our method,we not only conduct an extensive set of experiments for approximate nearest neighbor search evaluation on multiple ANN datasets,but also also validates the merit of our algorithm on the performance of semantic image segmentation based on label transfer.to obtain accurate segmentation result,we take advantage of jointly segment-ing similar object,and To ensure that our processing image group contain similar object,is able to co-segment image set with noisy via measuring the attention score of objects on each image.our algorithm Secondly,we propose a co-segmentation algorithm on an image set.To improve the robustness of segmentation algorithm,we introduce a metric called Attentiveness.Comparing to other methods,the pro-posed algorithm is robust to do co-segmentation on noisy image set by making use of this measure.To further separate images into the group of more similar objects,we employ a subcategory cluster on the image set.With the perfect image set,we introduce an co-segmentation algorithm via combining local shape prior and global shape prior.By solving an energy minimization problem,we further improve our segmentation result via iteratively updating the local and global prior with the intermediate segmentation.The empirical evaluation shows that the proposed al-gorithm is robust on general image group and obtains the improved segmentation compared to previous approaches.Finally,we introduce a very fast algorithm for image semantic segmentation based on deep learning technique.Although many methods based upon Convolu-tional Neural Network have shown remarkable segmentation results,they are still suffering from slow process speed due to the expensive computation cost of large network architecture.We avoid this by building our algorithm upon small network and combining two novel blocks to improve the segmentation.Our first proposed block called Hierarchical Dilation Block consists of multiple level structure of dila-tion convolutional layer,which is able to directly utilize multi-scale feature.The second block called Coarse-to-Fine Block is introduced to refine feature that be-comes coarse and loses spatial detail due to the pooling operation of network.This block employs bypass structure and auxiliary loss to archive this.The bench-mark evaluation show that our method runs two times faster than current fastest algorithm and importantly obtains better segmentation result.