Vehicle Recognition Combining ResNeXt Network and Label Smoothing Regularization

The commonly used cross-entropy loss function only calculates the loss for the correct position of the label, but ignores the influence of the wrong label. This makes the model too confident, resulting in lower prediction accuracy during model testing than training. We combine the ResNeXt network model with the label smoothing regularization method. Our model for fine-grained recognition of vehicle can improve the accuracy of vehicle recognition.