Target Classification For Complex Scenarios

Target classification is a core problem in the field of computer vision.In recent years,with the development of deep learning technologies,remarkable progress has been made in target classification.However,target classification techniques are mostly based on images.Unfortunately,images are difficult to obtain under poor network conditions and occluded environment.Therefore,it is very important to study how to classify objects when the associated images are unavailable.In this paper,we give a preliminary study under two difficult image acquisition conditions: 1)under poor network,i.e.,when images are not conducive to transmission;and 2)with poor illumination or occlusion.For the first scenario,this paper proposes to employ user click data,a kind of text data type that can be efficiently transmitted in the network,to construct sematic feature for images;while for the second case,we proposes to use the scattered field data for robust image representation,since scattered field will not be easily interfered by visible light.Consequently,we mainly focus on improving classification performance of click data and scattering field data,and our work can be summarized as follows:(1)Target classification method based on deep click feature.Click data contains valuable semantics,which are consistent with human perception.The click data records the user click count for each image-query pair.With user click data,a target can be represented as a query click vector with a pre-defined query space.However,this vector is highly sparse and has limited representational ability.In this paper,we divide the query text into words via natural language techniques,and utilize TF-IDF algorithm to represent the object by a click feature vector.Particularly,considering the part of speech for each word item,we construct several word dictionaries of different parts of speech.Afterwards,we generate a set TF-IDF click feature sub-vectors for each image for final classification.The experimental results show that: 1)the click feature vector with different part of speech has higher classification accuracy than the unified click feature vector on the whole dictionary without considering part of speech.The click feature tensor has rich semantics and can achieve better classification performance.(2)Target classification method based on electromagnetic scattering feature.Scattered field data is a kind of data received by electromagnetic sensors after detecting objects by microwave.Scattered field data is a kind of complex data,which is quite different from the real data and not be efficiently addressed in traditional real-valued networks.Therefore,we propose a complex convolutional neural network composed of complex components with attention mechanism to address this issue.In addition,we investigate the model robustness by testing the performance under heavy noise and limited aperture,which are very common in practice.The experimental results show that the complex convolutional neural network with attentional mechanism helps effectively learn the deep feature of the scattered field data and has good robustness.