The naming of rock hand specimens is usually conducted by geological workers based on observed mineral composition, texture characteristics, etc., combined with their own knowledge reserves. The accuracy of identiffcation results is limited by the experience, research interests, and identiffcation level of the identiffer, as well as the complexity of the rock composition. To improve the efffciency of rock hand specimen identification, this paper proposes a method for rock image recognition and classification based on deep learning and the Inception-v3 model. It encompasses the preprocessing of collected photographs of typical intrusive rock hand specimens, along with augmenting the sample size through data augmentation methods, culminating in a comprehensive dataset comprising 12501 samples. Experimental results show that the model has good learning ability when there is sufffcient data. Through iterative training of the Inception-v3 model on the rock dataset, the accuracy of rock image recognition reaches 92.83%, with a loss of only 0.2156. Currently, several common types of intrusive rocks can be identified: gabbro, granite, diorite, peridotite, granodiorite, diabase, and granite porphyry. Software is developed for open use by geological workers to improve work efffciency.