Baseline drift occurs universally in many sorts of analytical chemistry data, such as in MALDI-TOF, Raman, infrared and XRD spectra. In the era of big data, automatic correcting methods are eagerly demanded. However, traditional baseline correction methods are impossible to execute fully automatically. They always depend on some preset parameters. To build parameter-free methods, utilizing current intelligent algorithms is the best choice. However, it is a great challenge to provide a huge number of labeled samples which are required for effective training. In this article, a novel strategy has been developed to train a deep neural network successfully for baseline correction. The impossible mission of preparing millions of manually processed training samples was avoided. Under the new scheme, the power of deep learning was freely applied to achieve straightforward full automation in baseline recognition. Numerical experiments on authentic datasets indicated that the new intelligent model outperformed traditional methods.