To address the issues of intense subjectivity and low reliability in sensory evaluation of umami intensity in solid composite seasonings, this study proposes a prediction model integrating near-infrared spectroscopy (NIRS) and deep learning. By screening 1963 commercial samples and optimizing data quality through consumer sensory evaluation capability assessment, one-dimensional convolutional neural network (1D-CNN) and two-dimensional convolutional neural network (2D-CNN) models were constructed for quantitative prediction. The results showed that without consumer screening, the model achieved a mean relative error (MRE) of 12.79%~15.86% and a correlation coefficient (R) of 0.70~0.74. After excluding data from 6 consumers with poor evaluation capability, the performance of the 2D-CNN model significantly improved (training set: MRE=4.94%, R=0.90; validation set: MRE=5.25%, R=0.87). This study demonstrates that consumer evaluation capability screening and 2D-CNN-based feature extraction effectively enhance prediction accuracy, providing a robust and objective technical solution for quality assessment and product development of solid composite seasonings.