Smoky Vehicle Detection Algorithm Based On Improved Transfer Learning

摘要

With the demand for pollution prevention has increased. Establishing a smoky vehicle inspection and evidence collection system will help to supervise the problem of excessive emissions of vehicles. Because the sample in the training data set is insufficient and unbalanced, the existing methods cannot obtain sufficient classification accuracy in the classification task of the smoky vehicle. This paper uses the transfer learning method to compare the effects of smoky car classification tasks on the five basic networks of DenseNet, NasNet, Inception, VGG19, MobileNet. In addition, for the classification task of the smoke vehicle, it is necessary to consider both the global information of the tail of the vehicle and the detailed information of the smoke area. In this paper, the global average pooling layer is changed to the spatial pyramid pooling layer, and it has achieved better results. Finally, the classification Precision nearly 0.1, Recall 0.3 are obtained, which is much better than the traditional classification method and the shallow non-transfer deep learning method.