Rock images recognition is a classic appilcation scenario of the perceptual intelligence represented by deep learning in geosciences. Previous studies on the identification of rock images have shown that the simpler architecture of deep convolutional neural networks model can achieve the same or higher accuracy than the complexer. It is a paradox which is contrary to the research revealed that increasing the depth of a model can improve its recognition accuracy on the ImageNet dataset. How to explain this phenomenon? The interpretability of the various ImageNet- pretrained models with plutonic rock images can give a explain to the phenomenon. Plutonic intrusive rocks are crystalline, euhedral—subhedral granular texture and massive structure. The classification of them is based on their mineral composition and relative content. Magmatic activity is extensive in Dabie Mountains , and Mesozoic plutonic intrusive rocks are widely exposed. The rock types include ultramafic intrusion, gabbroid, dioritoid, syenitoid and granitoid, which basically cover the rock types in the classification scheme of plutonic intrusive rocks recommended by IUGS. The plutonic intrusive rock images were collected from the Dabie Mountains and were used to train using various pre- trained deep CNN. Using the local interpretable model—agnostic explanation (LIME) and the convnet visualization, the visual explanations for the feature extraction operation were given. Our experiments demonstrated that the simple struture convolutional neural networks model also can extract the color features and the texture features. For deep learing of rock images, the network structure is not always the deeper the better.