Rail surface state recognition method based on improved metric learning under small sample

The invention discloses a rail surface state recognition method based on improved metric learning under a small sample, and the main body of the method comprises five stages: a data collection and processing stage, a multi-scale feature extraction stage, a feature splicing stage, a measurement stage and a result display stage. A pyramid splitting attention mechanism is introduced in the multi-scale feature extraction stage, spatial information of different scales is captured, and the model recognition precision and the training speed are improved; in the feature splicing stage, a feature splicing module is started, a deep local splicing character is introduced, local descriptors of a feature graph are spliced, the influence of irrelevant information such as a background is reduced, and meanwhile local features with obvious discrimination are reserved; in the designed measurement stage, a convolutional neural network is used for replacing a fixed measurement formula, and fitting measurement of a combined featu