Abstract: Traditional manual rock mass classification exhibits various disadvantages, such as strong subjectivity, poor real-time capability, and reliance on geological engineers’ experience. To address these challenges, a rock mass classification system based on the BM-Inception network is proposed to realize accurate and efficient classification during full-face tunnel boring machine (TBM) construction. The morphological features of the muck—particle size, particle size distribution, and edge sharpness—are collected to analyze muck images and determine rock mass grades. First, an extreme value difference pooling layer is designed to enhance muck boundary features. Based on a custom activation function for noise suppression, the filtered boundary features are adaptively fused with global features. Second, a multiscale feature fusion module is developed. Cascaded small convolutional kernels and extra feature extraction branches are adopted to improve the muck feature extraction and simultaneously reduce the computational complexity of the system. Finally, an engineering-level real-time classification system with a client-server architecture is established, and a grouping strategy is introduced to boost classification accuracy. The proposed system achieves a single muck image classification accuracy of 89.29% on a self-built dataset. The system is deployed at two TBM construction sites of a plateau tunnel, where conducting tests on a total of 301 muck image groups result in an average group classification accuracy of 91.69% and an average analysis time of 041 s per group. The proposed system fully satisfies the requirements for rock mass classification in practical TBM construction.