Research On Multi-attribute Classification Of Human Images For Mobile Applications

Multi-attribute classification of human images is an important research topic in computer vision.It belongs to a multi-label classification,which is mainly used to classifies various attributes of human beings.Compared to the single-label classification task,the multi-label classification task is more complicated and requires solving many problems at the same time,so it is more challenging.In addition,due to the limited computing power of the mobile phone,in order to deploy the model to the mobile terminal,the method of model compression is also required to ensure the real-time performance of the entire system.In order to solve these two problems,this thesis starts from two aspects,aiming to improve the accuracy of the model while ensuring the adaptability of the mobile platform of the model: on the one hand,the accuracy of the model is improved through a complete two-level consistency matching mechanism;On the one hand,the knowledge distillation algorithm based on the batch normalization layer reduces the size of the model while ensuring the accuracy of the model.Finally,this thesis implements a compliance judgment system based on deep learning.The main research contents of this article are as follows:(1)Research on human image multi-attribute classification algorithm based on attention mechanism: The matching mechanism of the visual attention map consistency needs to manually design the mapping relationship between attention maps,which is difficult to design.Aiming at this problem,this thesis proposes a two-level simplified consistency matching mechanism,which does not need to manually design the complex mapping relationship between attention maps,which reduces the design difficulty of consistency matching.After that,we proposed a complete two-level simplified consistency matching mechanism.On ResNet50,the mAP was increased from 86.8% to 87.1%.Considering that on a large network such as ResNet100,the mAP of visual attention consistency algorithm can only reach 87.5%,so we think this algorithm is simple and effective.(2)Research on knowledge distillation algorithm: At present,the mobile terminal has limited computing power,so for deployment considerations,the model needs to be compressed.This thesis proposes a distillation method based on the BN layer.The parameters in the BN layer in the teacher network are directly passed as knowledge to the student network.Under the distillation framework with MobileNetV3-large as the teacher network and MobileNetV3-small as the student network,this distillation method can reduce the mAP gap between the student network and the teacher network to 0.33%.Later,this thesis designed a more streamlined MobileNetV3-small-small student network structure.With the help of our proposed BN-layer-based distillation method,the mAP gap between this student network and the teacher network is only 1.01%,and successfully reduced the model parameter amount to 22% of the teacher network,and reduced the model calculation amount to 15% of the teacher network.(3)Implementation of compliance judgment system: Based on the two algorithms mentioned above,this thesis implements a lightweight compliance judgment system with excellent results.In order to pursue better model performance,in the compression framework,we did not use the MobileNetV3-small-small network,but instead used the MobileNetV3-small network as the student network.In the end,we successfully deployed the model to mobile phones.