Image Semantic Segmentation Based On Region Proposal Network

In the field of digital image processing,image segmentation technology has always been a hotspot and difficulty in many research directions.The research work on image segmentation technology started in the 1960 s.After years of development,algorithms for image segmentation have emerged in an endless stream.This type of image segmentation algorithm is usually based on color,texture or shape of the underlying information for image segmentation,often ignore the spatial information and semantic information in the image,so only on certain types of images has a good effect and can not be applied to All the images.Aiming at these problems,an image semantic segmentation algorithm based on regional proposal network is proposed.This algorithm combines two kinds of deep convolutional neural networks,which is a regional proposal network and a full convolution segmentation network,to achieve better preservation of image spatial information and extraction of semantic information.The full convolutional segmentation network takes as input pixel-level images,uses convolutional layers to extract image features,and uses the upsampling method to restore the extracted features to the original image size.Aiming at the shortage of training data for pixel level annotation in the process of training and testing of full convolutional segmentation network and the problem of rough segmentation and lack of detail in image segmentation,the regional proposal network is introduced into the existing network framework to implement a joint network.The model is RPN-SegNet.Using the regional suggestion box generated by the regional suggestion network to include category marker information can effectively enhance the image segmentation performance of the full convolutional network segmentation model and improve the final segmentation effect of the image.At the same time,a method of weakly supervised learning can be implemented to make up for the shortage of training data.In addition,the binarization network algorithm is introduced into the RPN-SegNet model framework.The binarization processing of the input data and the weight filter at the network layer is used to solve the problems of low convolutional neural network operating efficiency and memory usage.Five-category and twenty-category image segmentation task experiments were conducted on the database PASCAL VOC2012 using fully convolutional segmentation network based on the regional proposal network.The experimental results show that the network model can effectively improve the classification accuracy of pixels by nearly five percentage points,and obtain a more precise segmentation effect.At the same time,the experimental results also show that the binarization network algorithm effectively improves the image segmentation efficiency of the RPN-SegNet network model and reduces the memory usage of the network model.