Research On Image Semantic Segmentation

Image semantic segmentation has aroused general concern due to its wide application and the development of compute vision.The traditional image semantic segmentation model was trained in pixels through the neural network,depth learning or other methods.Although the model adds local features to the pixel features,it ignores the integrity of the image objects and transforms the recognition of the image objects into pixel recognition.The recognition process of the model is too machine-type,which leads to the loss of image objects and the lack of contours of image objects in semantic segmentation results,and the accuracy of the final semantic segmentation results is not enough for further image processing.To overcome these shortcomings,we imitated the biological vision to recognize the image object.Firstly,we learned the holistic of objects in one image,then identified these objects using the learned holistic.According to the idea above,we proposed a semantic segmentation based on the image object candidate region.Based on the above ideas,the main work of this paper is as follows:(1)In view of the traditional image segmentation technique has a good expression effect,a novel semantic segmentation method based on the traditional segmentation method was proposed in this paper.The segmentation blocks were generated and would be used as the unlabeled object candidates.The proposed segmentation model first exploited the Structural Forest to generate the pixel-based edge probability map,and then generated the initial segmentation blocks with the improved watershed algorithm.In order to improve the overall-describe ability of the blocks,and avoid excessive segmentation and reduce the training expenses,an appropriate threshold was selected to merge redundant blocks,which also resulted in optimizing the contours of the segmented and less computation cost.In this way,we obtained high accurate contour of the image object.In the last step,pre-trained stochastic forest and support vector machine model were occupied in the semantic segmentation;finally results with good contour information were obtained.(2)As forecasted on a single scale,image semantic segmentation method based on the block of image segmentation cannot describe the object contour well,an image semantic segmentation method based on image hierarchical tree is proposed.In this method,the structural forest method is used to generate the contour model.And then the multi-scale contour map is produced from the contour model by Ultrametric Contour Map algorithm.To achieve hierarchical tree,the support vector machine is adopted to train multi-scale contour map.Finally,the conditional random forest is applied to get the final result of image semantic segmentation.Experimental results show that the proposed image semantic segmentation method improves the accuracy,stability and rate of image semantic segmentation effectively in complex outdoor environment.