Visualization of Railway Scene Classification Model via Grad-CAM
Bing Zhao, Ping Li, Mingrui Dai · 2018
Railway detection task produces a large number of images, the lack of effective image classification method makes it difficult to analyze detection image deeply. Using convolutional neural networks (CNN) to realize railway image scene classification is an effective technical means. This paper propose a method to reduce the bias of database by Gradient-weighted Class Activation Mapping(Grad-CAM) to effectively improve scene classification accuracy, and achieve accuracy of 95.3%(top3) on Railway12 database. Our approach combines two insights: (1) Small quantity of railway scene database make it hard for CNN to achieve high performance, transfer pre-trained ImageNet-CNN to fine-tune on railway scene database (2) Introduce Grad-CAM visualization method to analyze model's classification pattern and intuitively show the possible bias of database, provide an intuitive way to reduce bias of dataset.