| With the rapid development of artificial intelligence technology,home service robots aimed at improving people’s quality of life have attracted widespread attention.However,the current level of intelligence of service robots is not high,and they lack the ability to understand and judge the target items,cannot distinguish between items and garbage,and do not have the ability to deal with garbage items by category.Aiming at the above problems,based on the human perception and understanding ability of objects,this thesis studies the improved YOLOv5 visual detection algorithm and the key technology of garbage sorting based on the visual understanding method based on knowledge graph.The main research contents include the following three parts:(1)Construction of a knowledge graph in the field of garbage.Aiming at how to characterize and store massive multi-source heterogeneous junk information knowledge,a knowledge graph in the garbage field is constructed.Firstly,the pattern layer and the data layer are divided by ontology definition modeling;then the knowledge triplet is obtained by multi-strategy knowledge extraction and knowledge fusion technology for data from different sources;finally,the knowledge triple is stored and characterized based on the Neo4 j graph database,and the structured representation and knowledge visualization of the knowledge in the garbage field are realized.Experimental results show that the knowledge graph in the garbage field can provide a knowledge basis for the determination of items,and the auxiliary detection algorithm can provide intelligent decision-making for garbage classification.(2)An improved YOLOv5 visual detection algorithm is proposed.Aiming at the problems of complex background and large morphological differences between targets in the home environment,which are easy to cause missed detection of targets,this thesis proposes an improved YOLOv5 visual detection algorithm.By introducing the parallel mixed-domain attention mechanism ESA module into the backbone network to enhance key feature extraction,and compared with the serial attention mechanism module CBAM,the results show that the introduction of ESA is better than the CBAM attention mechanism module.Secondly,the localization regression loss function and NMS non-maximum suppression are replaced and optimized,and CIo U Loss is replaced by GIo U Loss as the localization regression loss function to improve the accuracy of localization regression;DIo U-NMS is used to replace NMS to improve the missed detection of target objects.The experimental results show that,compared with the original YOLOv5 detection algorithm,the improved visual detection algorithm improves the recall rate by 7.2%,the precision rate by nearly 7.9%,and the m AP by 5%.(3)This thesis proposes a visual understanding garbage classification method based on knowledge graph.Aiming at the problem that existing detection methods cannot distinguish between items and garbage,a visual understanding garbage classification method based on knowledge graph is proposed.By improving the YOLOv5 visual detection algorithm and OCR technology,the key features of item entities,relationships and attributes are obtained.Based on the knowledge map of the garbage domain,semantic matching is performed using multiple templates to determine whether the detected items are garbage and what kind of garbage is it,and realize the goal of intelligent garbage classification. |
