The Research On Object Attribute Recognition Method Based On Convolutional Neural Network

Object attribute recognition is one of the core issues in computer vision and its technology can be applied in the field of smart city transportation.The explosive growth of the number of vehicles put tremendous pressure on urban traffic,and the smart city transportation system can effectively alleviate traffic pressure.Accurate object attribute recognition can improve the performance of vehicle control,traffic monitoring and vehicle management in smart city transportation systems.With the rapid development of big data and artificial intelligence technology,deep learning has been widely used in object attribute recognition tasks.Compared with the traditional pattern recognition methods,the convolutional neural network based on deep learning can automatically extract image features,which significantly improves the accuracy of target attribute recognition.Therefore,the method of convolutional neural network for object attribute recognition has important application value.In this paper,the object attribute recognition is divided into two coupled sub-problems:object location and object classification.For the normal objects,a cascading model combining the object detection algorithm and the convolutional neural network is established based on the online hard example mining strategy.For the small size objects,the network structure of the cascading model is improved by network structure compression and multi-level feature map fusion methods.For the hard-detected objects,the cascading model is improved by the image enhancement algorithms.The paper is divided into four parts as follows:(1)For the object attribute recognition problem,an object location and recognition cascading model based on the object detection algorithm is established by combining Faster R-CNN and CNN in order to reduce the interference of background information for the recognition model.Experimental results show that the accuracies of CS-CNN model are increased by 22.91%,22.01%and 6.33%compared with CNN,Fast R-CNN and Faster R-CNN,respectively.(2)For the problem of normal object attribute recognition,an object location and recognition cascade model is established by designing a two-stage improvement strategy.In the location part of CS-CNN,the Faster R-CNN based on the online hard example mining strategy is designed to compensate for the defect of positive and negative sample imbalance.In the classification part of CS-CNN,PReLU and Focal Loss are introduced to improve the activation function and loss function in order to increase the accuracy of attribute recognition.Experimental results show that the accuracies of CS-CNN with both improved activation and loss function are increased by 6.82%,4.59%and 2.07%compared with original CS-CNN,CS-CNN of improved activation function and CS-CNN of improved loss function,respectively.(3)For the problem of small size object attribute recognition,the network structure of CS-CNN is improved by network structure compression and multi-level feature map fusion methods in order to increase the recognition accuracy of CS-CNN for small size object Experimental results show that the accuracies of the CS-CNN with improved network structure are increased by 32.25%?23.44%?19.29%and 17.21%compared with CNN,Fast R-CNN,Faster R-CNN and original CS-CNN,respectively.(4)For the problem of hard-detected object attribute recognition,the characteristics of hard-detected objects in blurred images are analyzed.The global histogram equalization and dark-channel dehazing algorithm are introduced to enhance the blurred images.The enhanced images are sent to the improved cascaded model CS-CNN to implement attribute recognition.Experimental results show that the accuracies of the improved CS-CNN by both the algorithm of global histogram equalization and dark channel dehazing are increased by 28.71%?6.15%and 4.35%compared with original CS-CNN,the improved CS-CNN by global histogram equalization and improved CS-CNN by the dark channel dehazing,respectively.