As an increasingly popular method in the machine learning field, deep learning is applied to industrial combustion processes in this work. Using easily available color flame images obtained by the charge-coupled device (CCD), a soft sensor system based on deep learning is proposed to predict the outlet oxygen content online. Unlike the traditional principal component analysis which only extracts linear features, a multilayer deep belief network (DBN) is designed to extract the nonlinear features for a better description of the important trends in a combustion process. With the DBN-based multilevel representation of the CCD flame images, more useful information about the physical properties of a flame can be characterized. Sequentially, in a supervised fine-tuning stage, two DBN-based regression models are simply constructed to obtain the nonlinear relationship between the flame images and the outlet oxygen content. The advantages of the proposed deep learning-based analyzing and modeling method are demonstrated via on-site tests in a real combustion system.