Key Technology Of Image Semantic Segmentation

Image semantic segmentation is a pixel-level image classification task,which has broad application prospects such as scene understanding,environmental perception and medical images etc.Traditional image segmentation methods mainly rely on combining hand-crafted features with some pre-defined rules,while they are difficult to deal with most complicated scenes.Based on deep learning,this paper will explore key technologies of image semantic segmentation and its applications in three different fields.The main works of this paper are as follows:The first one is natural image semantic segmentation.Aiming to treat with several problems in common such as mismatched semantic relationship,confusing and tiny categories etc.,we firstly systematically analyze the existing semantic segmentation frameworks,and then implement and compare three representative deep segmentation networks on public PASCAL VOC dataset,and obtain heuristic conclusions.Secondly,an attention-based brain MRI tumor segmentation is proposed.To address the problems of uncertain size and shape of brain tumors,this paper proposes Attention Network(AN)and Spatial Recurrent Encoder(SRE)structure.AN is designed to localize latent tumor regions.SRE spatially samples the features from the convolutional feature maps to encode contextual information.The experimental results on BRATS 2015 dataset demonstrate the effectiveness of the proposed algorithm.Thirdly,an effective cancer detection algorithm for breast pathological Whole Slice Imaging(WSI)is proposed.For the super-resolution WSI with(200,000 X 1000,000)pixels,a patch-based detection framework is presented.The framework consists of three modules:image pre-processing,patch classification and post-processing.In the pre-processing stage,an improved Ostu algorithm is applied to separate the background,then CNN classifier is trained for patch classification.In the post-processing stage,morphological and statistical features are extracted from tumor heat-map,and a random forest classifier is learned to determine the final results.The experimental results on CAMELYON16 dataset indicate that compared with the state-of-the-art methods,the proposed algorithm achieves a comparative performance.