| Warehousing management is a very important link in the assembly and maintenance line of high-speed trains.It can greatly reduce the manpower and material resources needed for warehouse operation by quickly obtaining the number and name of materials in the process of storing and retrieving materials.Large-scale object recognition and retrieval are mainly based on the image captured by the camera to obtain the information of the type and serial number of parts.By analyzing the characteristics of spare parts in the actual application scenario,this paper studied the methods of object recognition and retrieval for intelligent storage,mainly as follows:In the actual industrial scene of motor vehicle maintenance,there are many types of spare parts and large intra-class gaps.At the same time,there are fine-grained categories that belong to the same large category but differ in details.This paper proposed a largescale object recognition algoithm based on multi-attention fusion,which used the attention mechanism of the SE(Squeeze-and-Excitation)module to fuse features at different levels to establish a multi-level structure.Firstly,we considered the fine-grained objects as a large category,and used the coarse-grained recognition network for largeclass recognition to reduce the impact of fine-grained objects on the recognition accuracy.Then,by using the attention mechanism,the feature channels that contribute more to classification were calculated for each large category that requires fine-grained classification,and the features of multiple levels were integrated as the biasis of finegrained classification task,so as to improve the accuracy of recognition.Experiments in this paper have proved that the algorithm can get better recognition results on the parts database constructed in this paper,with the recognition accuracy rate of 97.84%,which was 6.68% higher than the single-stage network.In order to solve the problem that the global features lack geometric invariance and the local features lack geometric space constraints,this paper proposed to retrieve objects using fusion of local features and global features.Besides,in this paper,a heuristic-based multi-layered object retrieval method was proposed,which could improve the speed and accuracy of retrieval by considering the prior knowledge.In this paper,the fusion method of local features and global features was proposed to overcome the shortcoming of simple global features lacking geometric invariance and simple local features lacking geometric space constraints,and improved the accuracy of retrieval.Through the experiments on the parts database,when the Top10 was returned as the result,the mAP value of fusion features was 0.967,which was improved compared with both local and global features.The heuristic-based multi-layered object retrieval method maked full use of the prior knowledge of the users,combined with the distribution characteristics of the database itself.It speeded up the retrieval speed,and improved the user experience of the system.The experimental results showed that the retrieval speed of multi-layer retrieval was 54% higher than that of single-layer retrieval when the accuracy of multi-layer retrieval was guaranteed.Relying on the key horizontal laboratory projects,this paper built a practical system,including an image acquisition module,a data pre-processing module,algorithms and system software implementation and analysis modules,to verify the effectiveness of the above theoretical algorithms.To sum up,this paper combined with the actual industrial scene of vehicle maintenance and repair,and aimed at intelligent storage,proposed a multi-attention fusion hierarchical large-scale object recognition algorithm and a heuristic-based multi-level large-scale object retrieval method,and verified the effectiveness of the two methods on the actual system. |
PDF Full Text Request