Rock identification is important for geoscientists and engineers in civil, mining, and hydrocarbon sectors because it contributes to geologic modeling, resource estimation, and stability examination. This study examines the performance of the three CNN architectures: convolution neural networks (CNNs), ResNet50, and VGG16 on granite, basalt, limestone, and marble rock representing the four geological rock classes. The models were trained on extensive thin-section microscopic images of rocks captured at 40× zoom using an advanced petrographical microscope. The classifiers had 26 textural features of thin-section images, mainly granularity, surface roughness, micro-texture, color gradients, hue variation, edges, grain alignment, and contrast between different regions. We employed traditional CNN, ResNet50, and VGG16 models on our diverse dataset with highly satisfactory results for rock classification. CNN with four convolutional layers achieved 92%, demonstrating basic feature extraction but having limitations in handling complex structures. On the other hand, transfer learning techniques mitigate the problem and secure 97% accuracy for ResNet50 and 98% for VGG16. VGG16, with its deep 16-layer architecture, proficiently captured minute details and preserved the highest accuracy. It is evident from the comparison that transfers learning approaches like ResNet50, and VGG16 are more effective at rock classification than the general CNN approach.