Weakly Supervised Image Semantic Segmentation Based On Local Region Growth And Faster R-CRR

With the development of science and technology,artificial intelligence has entered our daily life and been widely used in the fields of autonomous driving,intelligent robots,medical image analysis and so on.Image semantic segmentation is a hotspot in the field of artificial intelligence,and its goal is to assign a label to every pixel in an image.At the moment,the further development of image semantic segmentation is restricted due to the lack of large-scale pixel-level labeled datasets.Therefore,researchers pay their attention to weakly supervised semantic segmentation,which replaces difficult-to-obtain pixel-level label with weak annotations to train semantic segmentation models.Image-level label is the simplest form of weak annotations,which is time-saving and is easy to obtain.In recent years,weakly supervised semantic segmentation methods based on the image-level label are emerging.However,image-level label only tells whether an object in an image.There is still a big gap between the segmentation based on image-level label and pixel-level labels.Currently,the image-level label based weakly supervised semantic segmentation faces many challenges:(1)Image-level labels are too rough to descript locations and boundary information of the object.Besides,the image-level labels cannot be directly used to train segmentation models.(2)It is difficult to fully distinguish the object and the complex background in the image,which is a common problem in the field of image semantic segmentation.(3)Most of the current deep learning based weakly supervised semantic segmentation models pay too much attention to the discriminative region of the object,which leads to incomplete segmentation of the object.To solve the above problems,this issue has tried a series of studies:(1)Due to the image-level label is lack of localization information of the object,we can use the Faster R-CNN to provide the bounding box with the help of the image-level label.However,the obtained bounding box is usually too vague to completely cover the object.Therefore,we provide a method to optimize the rough bounding box.First,an image is over-segmented into many superpixels.For irregular and small superpixels,bilinear interpolation algorithm is used to transform the superpixels into a uniform size,and the corresponding histogram of oriented gradient features are extracted.Then,a region adjacency graph is established,where a vertex represents a superpixel.Subsequently,the breadth firstsearch algorithm is utilized to traverse the adjacency matrix,which is aim to determine whether the superpixels belong to the object.An optimized bounding box is achieved with the help of the labeled superpixels.Finally,image semantic segmentation is completed using the optimized bounding box and the Grabcut algorithm.Experimental results show that our method performs better than some methods on PASCAL VOC 2012 and MRSC-21 datasets.(2)Most of the current deep learning based weakly supervised semantic segmentation models pay too much attention to the discriminative region of the object,which may lead to poor semantic segmentation results.Thus,we present the MDCDSRG(Multi-Dilated Convolutional Deep Seed Region Growing,MDCDSRG)method.Specifically,to overcome the disadvantage of the DSRG(Deep Seed Region Growing,DSRG)algorithm,we introduce multiple-dilated convolution blocks to augment multi-label classification network,and obtain multi-scale object localization maps.Then,the obtained multiple object localization maps are aggregated into one localization map,i.e.seed cues.The seed cues contain discriminative and surrounding regions of the target,which are larger and denser than the cues generated by other algorithms.Therefore,it can overcome the problem that the initial seed cues are always small and sparse.Finally,the image semantic segmentation is accomplished by feeding the acquired seed cues to the deep seeded region growing algorithm.Experimental results show that the segmentation accuracy of our method is 62.7% on PASCAL VOC 2012 dataset,which outperforms recent reported methods.