Abstract: Marbling is an important index that affects the quality of beef. However, the evaluation of marbling in beef processing enterprises is operated by professional, who test the beef by comparing samples with standard images, which is subjective and susceptible to the environment. To solve these problems, this study proposed an intelligent classification method based on deep learning and designed a neural network with 4-layer convolution including input layer, convolutional layer, pooling layer, fully connected layer and output layer. The automatic extraction of marbling features was achieved, and a beef marbling detection application based on mobile phone was developed. In this study, 1 800 images were divided into calibration set, validation set and test set according to the ratio of 3:1:1. The training process of deep learning model depends on a large number of labeled samples, and the training effect of a small amount of data is poor. Therefore, methods like rotating, mirroring, adjusting brightness, contrast, and increasing noise were used to amplify the sample data in this study. Finally, 5 400 images of calibration set and 1 800 images of validation set are obtained. The calibration set was used to train and adjust the internal parameters of the network, and the validation set was used to test the model. In order to further explore the accuracy of the model, this method was compared with the traditional machine vision method. And the number of large fat particles, medium fat particles, small fat particles, the total area of fat and the evenness degree of fat distribution were calculated. According to the above characteristics, a multiple linear regression model was established to identify the grades. The results showed that the method used in this paper had good classification ability for marbling, and the detection accuracy of validation set was 97.67%, which was higher than the traditional machine vision method. Samples with error grade did not span two levels. Through the observation of the misjudged images ，the marbling richness of misjudged images was similar to that of misjudged images, and the marbling score was between the labeled grade and the misjudged grade. Finally, an application program for marbling detecting was written. In this application, you call use the model generated in this study to realize real-time detection for marbling. Then the application was used to recognition the samples of test set. The accuracy was 95.56%, and the recognition time was less than 0.5 seconds per image. In this paper, a beef marbling mobile phone evaluation system was developed based on the advantages of convolution neural network with high classification accuracy and fast speed of smart phones. This method could improve the detection efficiency of beef marbling. Compared with traditional detection equipment, the smart phone as a handheld collection terminal has the advantages of small size and high efficiency. The popularization of smart phones provides a broad development space for this method. Compared with ARM portable devices, this method uses mobile phones as the detection carrier, which saves the need to purchase additional auxiliary equipment and reduces the cost of hardware development.