Being one of the most fundamental embranchments of deep learning theory, scene categorization technology has been extensively researched because of its great value in engineering application, especially in the field of remote monitoring and intelligent fault detection. To bridge the gap between theoretical accuracy and practical performance of relevant classification models which is mainly caused by nonstandard labeling information, this paper builds a normative dataset composed of 10,000 high-quality manual labeled images from the power sector, and proposes a high-performance multi-label classification model utilizing deep convolutional neural network (CNN) inspired by lnception-v4 [1] on this basis. Experiments demonstrate that the model proposed achieves an accuracy of 94.125% on the test set and thus can be deployed into practical intelligent surveillance scenarios.
展开▼