Research And Application On Surveillance-Oriented Appearance Charcteristics Recognition

Directed by Xi,the Chinese government has been putting great effort into the industry of artificial intelligence,and the police department is one of its core application fields.Considering that deep convolutional neural networks have achieved outstanding performance in the task of image feature learning and object classification,our research focused on employing and improving deep convolutional neural networks in age estimation,gender classification,and ethics classification tasks under police-orientated scenarios.The challenges we are facing include accuracy and false positive rate improvement,resource restrictions,as well as heterogeneous environment adaption.To conserve police force,not only excellent accuracy but a low false positive rate is expected.And because police equipment normally has limited computation power and are required in various situations,the model must be adaptive,efficient and restrained in scale.These issues have addressed as follows:(1)In this thesis,three network models were introduced for gender classification,ethnicity classification,and age estimation tasks in order to reduce the computational cost,improve accuracy and avoid false alarm.Proposed a concise network structure to reduce resource usage in gender and ethnicity classification,while combined VGG convolutional layers and regression network in age estimation for best accuracy.(2)Developed an alternate adversarial learning algorithm to transfer learned model for different conditions and solve the heterogeneous environment adaption problem.It is common for adversarial learning to collapse due to imbalance players,and the learning algorithm we developed deals with this by combing label smoothing,noise addition,activation function swapping and weight updating modification using parameters obtained through experiments.(3)The proposed adversarial learning algorithm was applied to our networks which is split into submodules.By connecting different modules,we were able to adapt model trained on labeled well-shot source images to target images closing the gap caused by heterogeneous datasets.