Aiming at the problem that liquid surface features are few and discrimination is low, which is difficult to be recognized and detected effectively by machine vision, two laser light sources with different wavelengths are used to irradiate liquid at the same time to improve the discrimination between different liquids. An automatic collection device of data set is designed to provide a large number of effective samples for model training and then a visual recognition model based on EfficientNetV2 deep neural network was constructed. After introducing cosine learning rate dacay into the model and regulating super-parameters to the best, the optimal method was formed to realize efficient training, and the prediction accuracy was further improved. The results showed that the visual detection system could obtain 100% prediction accuracy, and successfully solved the problem of few features in liquid visual detection.