The artificial intelligence approach in the monitoring of laser welding is widely studied to ensure the good weld quality. In this study, a deep neural network (DNN)-based monitoring framework for weld shape prediction was established. This weld shape prediction model takes an optical signal as its input and generates a weld shape image as its output. Optical signals were acquired using the optical emission spectroscopy (OES). For the training of prediction model, this acquired optical signal was preprocessed using three methods: principal component analysis (PCA), spectral feature extraction (SFE), and relative intensity calculation (RIC). The weld shape image used for model training was generated by obtaining an eigen-image using the singular value decomposition (SVD) method. Supervised learning was used, and the generated images were labeled with optical signals according to the process conditions. Model evaluation was performed by comparing the weld area in the eigen-images with the one in predicted images. The two weld images were completely overlapped using feature points (e.g., top surface or tip of welds), and then, image accuracy was quantitatively determined by dividing the intersection area of these images by their union area. The weld shape predicted by the DNN model showed the accuracy of ~90% for all input data types.