Snow Scene Segmentation Using CNN-Based Approach With Transfer Learning

Images from CCTV cameras can be used for analyzing disaster situations in a particular area. Snowfall is one of the weather conditions that could cause natural disasters in Japan. It is possible for a machine to detect snow and mark these areas in that image. There are existing convolutional neural network-based (CNN-based) frameworks that can achieve high accuracy in an object classification task. However, these frameworks cannot define or mark the affected area then display the result. To address this problem, this paper proposes a method to develop a model using CNN frameworks, with the transfer learning technique. We use transfer learning to reduce training time and computing resources while maintaining high performance in a snow detection task. For transfer learning, the pre-trained weights from the VGG19 dataset is used. In this work, we use images from CCTV cameras, which were obtained from a publicly accessible website in Japan. After evaluating the model we performed post-processing on the data to further reduce the error in the results. Our proposed method can achieve an average sensitivity of 80.90% and specificity 98.54% for snow detection in real-world images.