Vehicle Detection in Satellite Images by Parallel Deep Convolutional Neural Networks

Abstract

Deep convolutional Neural Networks (DNN) is the state-of-the-art machine learning method. It has been used in many recognition tasks including handwritten digits, Chinese words and traffic signs, etc. However, training and test DNN are time-consuming tasks. In practical vehicle detection application, both speed and accuracy are required. So increasing the speeds of DNN while keeping its high accuracy has significant meaning for many recognition and detection applications. We introduce parallel branches into the DNN. The maps of the layers of DNN are divided into several parallel branches, each branch has the same number of maps. There are not direct connections between different branches. Our parallel DNN (PNN) keeps the same structure and dimensions of the DNN, reducing the total number of connections between maps. The more number of branches we divide, the more swift the speed of the PNN is, the conventional DNN becomes a special form of PNN which has only one branch. Experiments on large vehicle database showed that the detection accuracy of PNN dropped slightly with the speed increasing. Even the fastest PNN (10 times faster than DNN), whose branch has only two maps, fully outperformed the traditional methods based on features (such as HOG, LBP). In fact, PNN provides a good solution way for compromising the speed and accuracy requirements in many applications.