Abstract: Rock classification and boundary determination of rock slopes are very important for the analysis of slope stability. At present, the artificial methods are inefficient and affected by subjective factors. So a convolution neural network model for the image set analysis of a rock slope is established based on Tensorflow. Through convolution operation and pooling operation, the feature information of 8000 original rock slope images is extracted and compressed respectively. Then the network model is trained to realize the automatic recognition and classification of the rock slope. The model is tested and analyzed by using the images of rock slopes in training set and testing set. The accuracy rate of the training set and the testing set is 98% and 90%, respectively. It is shown that the network model after training has good robustness and achieves ideal training effect. Next, the color of different rocks on the slope is taken as the main basis. The boundary of different types of rock on the rock slope is calibrated by the deep learning boundary extraction technology. To verify the effectiveness of the algorithm, the standard color image of the rock slope is selected for simulation experiment, and the results of boundary detection are accurate. The network model established by deep learning realizes the requirements of rapid and automatic rock identification and boundary range division of rock slopes, and introduces the rock slope information acquired by image recognition into the GeoSMA-3D software independently developed by the team,as an important parameter for determining the grade of rock slopes.