Semantic Segmentation Based On Wavelet Convolutional Neural Network

Image semantic segmentation is the core task of computer vision.Its basic idea is to assign a label to each pixel in the image according to the category of the object to which the pixel belongs.Image semantic segmentation can help computers better understand the content of the image,so it has a wide range of applications in the field of artificial intelligence such as robots and unmanned driving.Although the semantic segmentation technology is becoming mature at present,the semantic segmentation of some special objects or scenes still needs to be further studied.For example,the semantic segmentation of transparent objects is very challenging due to its transparent characteristics.In recent years,convolutional neural networks have been widely used in the field of image semantic segmentation,and a lot of excellent research results have been produced in this field.However,due to the information loss in the down sampling layer of convolutional neural network,the precision of semantic segmentation will be affected,especially for transparent objects which require high precision segmentation.In addition,convolutional neural networks extract local information of images through the convolutional layer.Although larger receptive fields can be obtained by deepening the network,there is always a lack of effective methods to extract global information in a single network layer.In order to solve these problems,a kind of wavelet convolution neural network is proposed in this paper.Firstly,aiming at the problem of information loss in the down sampling layer of convolutional neural network,a down sampling module based on wavelet lifting is proposed to save more useful information.Then,aiming at the difficulty of obtaining global information in convolutional neural network,a self-attention module based on context information is proposed,which fully considers the local similarity of natural images while obtaining global information.Finally,in the semantic segmentation task of transparent objects,we make experimental comparisons with other mainstream segmentation networks in terms of segmentation accuracy and global information extraction methods,and verify the effectiveness of our proposed model.