RockNet for Classification of Rock Slice Image

Classification of Rock Slice Image is one of the basic means of studying the environment in the geological industry, which is of great significance for the development of the geological industry. This paper proposes an approach, RockFPN, which accurately captures the location of regions of interest, maintains the input feature map channels dimensionality, and combines deep high-level semantic information with shallow lowlevel spatial details. This design enables each hierarchical feature map to retain rich semantic information simultaneously. Based on the classical ResNet backbone, the proposed RockFPN achieves an accuracy of 98.8 % on the photomicrograph dataset of rocks for petrology teaching at Nanjing University, with an inference speed of 17.44 ms per image on an NVIDIA GeForce RTX 4060 These results demonstrate its potential for applications. Code and datasets are available at https://gitee.com/zang-hu/rock-net.