Long-tailed distributions are prevalent in visual recognition tasks, where a few head classes have abundant samples while most tail classes suffer from severe data scarcity, significantly limiting model generalization on tail classes. Traditional approaches such as resampling and reweighting partially alleviate this issue but often overlook the intrinsic semantic structures within classes, hindering effective alignment between head and tail distributions. To address these challenges, this paper proposes an Intermediate Class-Driven Balance Adjustment Method (IC-BAM). It adaptively subdivides head classes into semantically consistent subclasses to enhance fine-grained discrimination, while employing adaptive generation strategies to augment tail class diversity. This collaborative strategy facilitates the progressive alignment of both head subclasses and tail samples toward intermediate classes, promoting structured balance across the entire class space. IC-BAM consists of two key modules: the Adaptive Hierarchical Subclass Clustering Module (AHSC) to improve head class discrimination, and the Adaptive Tail Amplification Module (ATAM) to strengthen tail class representations. To accommodate the two-level class hierarchy, a customized loss function and an evaluation metric for optimizing intermediate class selection are designed. Extensive experiments on CIFAR10-LT, CIFAR100-LT, and a real-world steel defect dataset validate the effectiveness of the proposed method, demonstrating significant improvements in overall and tail class classification performance.