Algorithm Research Of Image Semantic Segmentation Based On Deep Learning

The task of image semantics segmentation is to classify all the pixels in the image,so that the image can be divided into several meaningful and interesting regions.It is one of the most critical links in the field of computer vision and the basis of image processing and analysis.Because of the challenge and importance of image semantic segmentation,the research of semantics segmentation algorithm has received great attention.In recent years,the high accuracy of deep learning in solving complex problems has made it widely used and studied in semantic segmentation tasks.Compared with the traditional image semantics segmentation algorithm,the image semantics segmentation algorithm based on deep convolutional neural network relies on the neural network to extract features,and it infers the segmented image with semantic information through the obtained features,so as to improve the accuracy of segmentation.The core task of this paper is to study the existing image semantics segmentation algorithm,analyze the advantages and disadvantages of the algorithm,propose an improved algorithm,and verify the accuracy and practicability of the improved algorithm by simulation.Firstly,in view of the time-consuming and labeling difficulties caused by strong supervision,the semantic segmentation of this paper trains deepconvolutional neural network with weakly-supervision,carries out object detection,and obtains the localization maps.The result is used as the segmentation mask of learning segmentation network to improve the training efficiency and the utilization of input data.Secondly,weakly-supervised objects detection brings incomplete information problem,the problem makes the positioning of objects of the same category easy to stick when performing objects detection,which makes weakly-supervised objects detection challenging.Aiming at the problems caused by weakly-supervised,the contribution of this article is a new architecture of two-stage cascaded network method is proposed.The localization maps obtained from the first stage network are used as input data to train the second stage network,so as to enhance the accuracy of localization maps.Third,although the multi-objective adhesion problem of localization maps has been solved,the above algorithm still has the issue of sparse localization maps.In this paper,an algorithm combining adversarial Erasing(AE)and two-stage cascaded deep convolutional neural network is proposed.In short words,on the basis of two-stage cascaded deep convolutional neural network algorithm,the regions located by the upper network network is erased as the input of the next stage network,so as to obtain a new localization regions.Repeating such adversarial erasing can localize increasingly discriminative regions diagnostic for image category until no more informative region left.Finally,the erased regions are merged to form a pixel-level semanticsegmentation mask that can be used for training a segmentation model.Finally,In order to improve the accuracy of edge contour location of semantics segmentation results obtained by learning semantics segmentation network,conditional Random Field(CRF)is added as post-processing operation in this paper.Experiments are conducted using The PASCAL VOC 2012 datasets,the results show that the classification accuracy of the two-stage cascaded deep neural network proposed in this paper is 87.2% under weakly-supervision,compared with the recent advanced algorithm,In the evaluation of object localization,the mIoU increases by 2.1% and the recall increases almost 9%.The semantics segmentation result of the two-stage cascaded deep convolutional neural network based on AE,which is post-processing by CRF,is 4.2% higher than those of the recent advanced algorithms.