Scene Recognition Based On Multi-task Learning

Scene recognition is the basis of robot environment understanding and behavior planning,and has great value for improving human life,it has been an important research topic of computer vision.Compared with other computer vision tasks,scene images have higher interclass similarity and intra-class differences,so it is more difficult to deal with scene recognition problems by general classification methods.Most of the methods of scene recognition focuses on the enhancement of local features to improve the scene recognition.These methods usually require local supervision information,and the implementation is more complicated.This thesis aims to propose a more efficient method for scene recognition by studying the correlation between scenes and objects.We innovatively use the multi-task learning method to deal with the scene recognition problem.We use a single network to simultaneously process scene recognition and object recognition tasks to improve the scene recognition performance,which is based on feature sharing and mutual promotion.First of all,this thesis studies the relationship between scene and object recognition tasks.Through the study on the datasets,the research methods and the relationship in CNN between scenes and objects,it is found that significant objects have an important influence on scene recognition,while scene recognition models have the ability of unsupervised object localization.We come to a conclusion that scenes are highly correlated with objects.Scene and object are different descriptions of the same phenomenon,and the simultaneous processing of the scene recognition and the object recognition is operable.Secondly,through the research of single-task scene recognition,the effectiveness of proposed SceneBlock for improving scene recognition is verified.And the SceneNet is designed to realize single-task scene recognition.SceneBlock is the basic unit of SceneNet,which is an innovation on short-cut structure of ResNet,by adding gate units to short-cut structure.At the same time,SceneNet preserves the positioning capabilities of the network through the Golbal Average Pooling(GAP)layer.The experiment results show that the proposed SceneBlock structure has the ability to extract richer features.The GAP layer is effective for maintaining the locating capability of network,and SceneNet has a significant improvement on scene recognition problem.Finally,the multi-task learning method is proposed to train the scene and object recognition model simultaneously,which is realized by SceneObj-CNN.SceneObj-CNN is based on SceneNet,which has the ability to extract richer features.Scenes and objects share the common middle and low-level features.The unsupervised object locating ability of scene recognition model enhances the object recognition effect,and the object category is used for local feature supervision to promote the scene recognition performance.The final experiment results show that the multi-task learning method has a significant effect on improving the scene recognition performance,and realize object recognition.The proposed networks structure and experimental methods also have reference significance for scene recognition and object recognition.