Real-Time UAV Vehicle Detection Based on Enhanced Feature Information

‍ ‍Vehicles from unmanned aerial vehicle （UAV） images were difficult to achieve accurate classification and location， due to the objects were small size and low-resolution. A light-weight feature extraction network was designed in this paper to provide multi-scale mid-/low-level feature that was integrated into the backbone network， which realized the transmission of mid-/low-level information. At the same time， the high-level semantic information was extracted from the backbone， which was beneficial to differentiate the target vehicle from background or other vehicle categories， then deep high-level semantic features and shallow features were fused to realize the transmission of high-level information. Thus， a similar bi-directional network was introduced that could effectively transfer information from different levels and enhance the feature representation for vehicles. Furthermore， multi-rate dilated convolution was proposed to obtain richer mid-/low-level information， and an effective feature fusion module was presented to integrate the mid-/low-level information into the backbone. The experimental results showed that the proposed algorithm could achieve accurate classification and location for UAV vehicles and realize the real-time application requirements.