To address the problem of imbalanced data distribution between healthy and faulty states in rolling bearings, we propose an end-to-end fault diagnosis framework that integrates Continuous Wavelet Transform with an Improved Deep Residual Network (CWT-IDRN). Compared to conventional residual networks, the IDRN introduces three major improvements. First, a spatial transformation module is employed to enhance image representation and emphasize regions of interest. Second, an attention mechanism is incorporated to amplify critical features and fuse them with residual sub-blocks. Third, a class-balanced reweighting strategy is proposed, alongside a novel loss function that replaces traditional logistic loss. Experimental results on three benchmark datasets demonstrate that the proposed CWT-IDRN outperforms existing intelligent diagnostic methods in both accuracy and generalization.