Research On Concealed Objects Detection Algorithm Of Multi-view Millimeter-wave Image Based On Deep Learning

Automatic detection of concealed objects in active millimeter-wave(AMMW)images is of great significance.It can not only avoid staff observing the human body AMMW images,which invades the privacy of the inspected persons,but also better prevent high rate of false detection and high rate of missed detection caused by human factors.At present,the concealed objects detection algorithm based on multi-view AMMW image makes use of the temporal sequence of AMMW images,so good results are obtained.However,research on how to make use of the temporal sequence is insufficient.Since it is difficult to obtain the AMMW human screening dataset,the dataset is generally small.how to make full use of the current small dataset and the temporal sequence of samples has become the key point of multi-view AMMW image concealed objects detection.In this thesis,a multi-view AMMW image target detection algorithm with good performance is proposed by combining the characteristics of AMMW images with the application of deep learning theory.The main work of this thesis includes:(1)By implementing the AMMW image concealed objects detection algorithm based on RetinaNet and LSTM,this thesis compares their performance,which proves the importance of temporal sequence in AMMW images.(2)This thesis uses the Gated Recurrent Unit(GRU)instead of LSTM to optimize the network structure of the model,which enhances the feature extraction ability of the model.Then,the effectiveness of the improvement was verified by experiments.(3)By adding ensemble learning method,the model has better generalization performance on small dataset.After above improvement,the multi-view AMMW image concealed objects detection algorithm based on GRU proposed in this thesis has achieved better detection results.