Image Semantic Segmentation Based On Region And Deep Residual Network

Computer vision technology is used the computer instead of human eyes to recognize,detect and track the object in images or videos,and then processes them in the next step.With the rapid development of deep learning,computer vision based on deep learning technology has gradually appeared in various areas of daily life.Image Semantic Segmentation is a branch of computer vision,more and more new applications require accurate and efficient segmentation mechanisms,such as autopilot,geographic information systems and medical image analysis.How to get accurate and efficient segmentation results is an urgent problem to be solved in the current semantic segmentation field.Semantic segmentation based on region uses multi-scale to extract overlapping regions,which can be capturing complete objects and canonical object parts.However,the process of this method is complex.The method based on full convolution network uses the self-learning feature of convolution neural network,which can be used for end-to-end training for pixel-by-pixel classification tasks,but this method usually results in inaccurate object boundaries.Aiming at the problems that some semantic segmentation algorithms have complex processes and rough segmentation results,we propose an end-to-end semantic segmentation model which combines region and deep residual network.Firstly,we use a region generation network to generate a set of candidate regions in the image.Candidate regions are described by bounding box,foreground mask,foreground size,and then stored as backup.Then using the deep residual network with dilated convolutions to get feature maps.Deep residual network was first used in image classification tasks and achieved good results,but the final output feature maps resolution of this network was low.In order to adapt to the task of image semantic segmentation,we proposed an extended convolution kernel to replace the common convolution kernel in the original residual network model,so that the network can output highresolution feature maps,which is convenient for the next step of segmentation and classification operation.Combining candidate regions and feature maps,we can obtain the feature of regions,and then map it to each pixel in the region.In order to highlight the prospect of the object,we proposed a method that combines region features with foreground features.Finally,using the global average pooling layer to classify images pixel-by-pixel.The global average pooling classification layer can receive input at any scale and classify it pixel-by-pixel.we also used multi-model fusion method,set different inputs in the same network for training to get multiple models,and then fuse features at the classification layer.We used mean method and voting method to get the final segmentation results.On SIFT FLOW and PASCAL Context datasets,the algorithm we proposed in this paper has higher mean accuracy than some state-of-the-art methods.Through the results of qualitative comparison experiment,it can be seen that our algorithm is accurate at the boundary of the object,and the segmentation edge is close to the ground truth labeling.Besides,our method also correctly identifies small objects.