Vehicle Attributes Recognition Using CNNs

Vehicle attributes are valuable clues of vehicle recognition,identification,retrieval and other related applications.It has many meaningful attributes,the most important of which is identification attributes such as vehicle license plate number,vehicle brand and vehicle color.Vehicle attributes recognition on surveillance vehicle images and videos not only can help people extract meaningful information from massive data,but also save data storage costs.In this dissertation,we mainly study on vehicle attributes analysis technologies for surveillance videos and bayonet images.By using Convolution Neural Network(CNN)technological framework,we study on fine-grained vehicle model and body color recognition.Also,a vehicle attributes recognition prototype system is designed and developed based on these technologies.The main contributions of the dissertation are as follows:First,a vehicle attributes recognition algorithm based on CNN is presented.By detecting the vehicle to locate the vehicle region,fine-grained vehicle types recognition is achieved by a CNN model.Experiments are performed on more than 100000 bayonet images,and the average accuracy reaches 99.3% with 250 vehicle classes.Second,a fast vehicle annotation algorithm is presented,which is based on CNN features clustering,indexing and retrieval.A small amount of vehicle samples was obtained after clustering by CNN features,which was used to create indexed data.Then,set the retrieval results as the final annotated results by retrieving unlabeled vehicles from the indexed data.Experiments are performed on more than 32000 bayonet images,and the average accuracy reaches above 90%,which minimizes the manual annotation cost.Third,a CNN-based multi-task learning algorithm is proposed.By designing architecture of CNN,we can recognize the two attributes of vehicle type and body color in only one CNN net.Experiments are performed on more than 42000 bayonet images,and the average accuracy of vehicle type and body color reaches 98.4% and 96.7% respectively.Meanwhile,comparing with processing two tasks separate,the algorithm can save half time.