A method for cleaning vehicle type database based on off-line and on-line clustering

Abstract

The invention discloses a method for cleaning vehicle type database based on off-line and on-line clustering. At first, the off-line vehicle type database is obtained by marking samples of various vehicle types, the training is carried out by using depth learn, and the output of the full connection layer of the training reverse second is taken as vehicle type characteristics. Secondly, all vehiclefeatures in each class are extracted for off-line clustering, and n class centers and corresponding thresholds are obtained. Then all vehicle types in the online vehicle type gallery are extracted periodically for clustering. The initial clustering centers are the n clustering centers obtained from the off-line clustering. After adding a random initialization center, the constrained clustering iscarried out to obtain the n+1 class. Finally, according to the off-line clustering threshold, the vehicle type data belonging to the first n classes are judged and cleaned in turn, and the data of the last class are cleaned. The invention can effectively delete the samples inputted into the warehouse by mistake under the condition of keeping the overall properties of various types of the on-linevehicle type warehouse unchanged, thereby maintaining the performance and stability of the long-term operation of the system.

Description

technical field [0001] The invention relates to a method for cleaning a vehicle model database based on off-line and on-line clustering. Background technique [0002] With the sharp increase in the number of motor vehicles, illegal and criminal vehicles are on the rise year by year, such as: hit-and-run, false vehicle licenses, vehicle license plates, motor vehicle speeding and other crimes are often staged. With the development of technology, the intelligent vehicle identification method is becoming a mature and effective means, which can be widely used in bayonet vehicle detection, license plate detection, vehicle retrieval and other aspects. [0003] In many applications, it is necessary to establish an online car model library. The car model recognition technology based on deep learning can achieve an accuracy rate of more than 98%, but in the long run, the continuous storage of wrong samples will still cause them to accumulate to a level where it is difficult to mainta...

AI Technical Summary

Problems solved by technology

[0004] Most of the existing database cleaning methods are general data cleaning or data cleaning for a specific field, and there is a lack of cleaning methods for vehicle model databases

Images