Vehicle recognition using multi-task cascaded network

Vehicle attribute recognition mainly contains two tasks: vehicle object location and vehicle category recognition. We propose a multi-task cascaded model MC-CNN, which integrates the improved Faster R-CNN and CNN. The first stage uses the improved Faster R-CNN network (IFR-CNN) to process the object location, and the second stage uses the improved CNN network (ICNN) to realize the object recognition. In IFR-CNN sub network, a max pooling and the deconvolution operation are added to the shallow layers of Faster R-CNN network. IFR-CNN can extract features from the different levels and increase the location information of shallow object. In ICNN sub network, we optimize the information extraction ability of high-level semantics in the middle layers and the deep layers of CNN network. The experimental results show that MC-CNN network proposed in this paper has better attribute recognition accuracy on BIT-Vehicle dataset and SYIT-Vehicle dataset than the single Faster R-CNN and CNN network models.