The present invention relates to a visual scene identification method based on a convolutional neural network. The method comprises the steps of: employing an original data set to obtain an image pairof perception deviation and perception variation; performing feature extraction of a pre-training convolutional neural network model (CNN); constructing a feature difference graph by employing features of different layers of the convolutional neural network; and constructing a new classification model to perform training of the feature difference graph.