Research On Key Technologies And Methods Of Image Semantic Segmentation Based On Deep Learning

Semantic Segmentation has always been a challenging task in the field of computer vision,playing an extremely important role in image understanding.The purpose of semantic image segmentation is to accurately segment the target object and assign a semantic label to each pixel in the image.In recent years,deep learning plays an extremely important role in computer vision.It has an efficient automatic feature extraction function,combining low-level features to form high-level features and obtaining the spatial correlation between different characteristics,which makes the deep learning algorithms have great advantages in extracting global feature information and local feature information of images.Based on these characteristics,deep learning also provides a new idea for image semantic segmentation.Although the feature of deep extraction of convolutional neural network are beneficial to image recognition,most of the pixel information is lost,which is not advantage to the end-to-end segmentation problem.Additionaly,the features of shallow layers retain most of the pixel information,which is not sufficient to object recognition.Based on these observations and the current research,this thesis mainly carries out the following research:(1)Deep Context Convolutional Network(DCCNet)has been designed,which combines the feature maps from different levels of network in a holistic manner for semantic segmentation.The segmentation outputs of DCCNets are post-processed using dense connected conditional random fields(CRF),where the relationship between the pixel categories in the image is taken into account,thereby further improving the segmentation performance of the image.(2)A novel encoder-decoder architecture has been presented,called dense deconvolution network(DDN),for semantic segmentation,where the feature maps of deeper convolutional layers are densely upsampled for the shallow deconvolution layers.The proposed DDN is trainable end-to-end,and allows us to fully investigate multi-scale context cues embedded in images.(3)A model compression algorithm has been proposed.Due to the excessive parameter size of the DCNNs model,the network occupies too much storage resources,which is not benefit to the application in mobile terminals and embedded chips.Based on the analysis,this paper proposes a low bit quantization algorithm to quantify the model parameters to achieve model miniaturization.The experimental results demonstrate that our method outperforms or is comparable to state-ofthe-art methods on PASCAL VOC 2012 or SIFTFlow semantic segmentation datasets.