Deep learning has gained popularity in the task of tool wear identification recently. As an important application of deep learning, however, there exist few public datasets and benchmarks for the research of visual identification of tool wear. To address this issue, we present a classification-based image dataset for carbide milling tool wear (NJUST-CCTD) and make it publicly available on the Github website. This dataset includes two categories: wear tools and no-wear tools. The two categories contain 5000 and 3000 photos, respectively. Based on this dataset, eight baselines are evaluated as references against this benchmark. To further improve the classification performance, we propose a novel cemented carbide milling tool wear intelligent classification framework (CMCNet). The framework consists of two modules: a deep learning based classification network and a multi-scale feature fusion denoising network called DSSNet. DSSNet is constructured with deeper network structure, connections across layers, and multi-scale sequence fusion module. It is capable of explicitly modeling the semantic and spatial correlation. Apart from DSSNet, the denoising module further improves the performance by adaptively altering the level of denoising based on the performance of the network. The two modules could be optimized with the backward gradient, yielding an end-to-end learning framework. On the basis of the dataset, CMCNet performed exceptionally well when categorizing photos intelligently. After 50 training epochs, the model outperformed the original classification network by 3.250%, achieving the top-1 accuracy of 95.375% on the test set. The NJUST-CCTD can be downloaded at https://github.com/paddy112233/PADDY