Machine learning (ML) provides an efficient method to predict the unknown properties during the exploration of new materials, but how to efficiently represent the molecules as input is still not fully solved. Inspired by image processing, one of the classical ML tasks, this work developed a method to predict the structure-dependent properties by converting the atom position into a three-dimensional (3D) molecular image and learning the structure features from the image via a classical convolutional neural networks. After trained with datasets larger than 12,000 species, a very high accuracy is obtained in predicting both theoretical molecular energy and experimental properties including melting points, boiling points, and flash points. Since stereoscopic information is explicitly and accurately represented by the molecular images, our model successfully distinguish the melting points and boiling points of molecules with similar structure, including those of trans-cis isomers.