China is the world's largest producer and consumer of fresh corn. Soluble solids content (SSC) is a key indicator of the quality of fresh corn, and there is an urgent need for effective and rapid non-destructive testing methods to respond to the market demand for the test. In order to realize the rapid and nondestructive detection of fresh corn SSC, a prediction model of fresh corn SSC based on near-infrared spectral features combined with a chemometrics method is proposed. Taking sweet corn as the research object, using the near-infrared (NIR) detection device built independently by the laboratory, we explored the multi-point acquisition method based on fresh corn materials to obtain the NIR diffuse reflectance spectra in the middle of the cob, and after the anomalous spectra were excluded by the Mahalanobis distance method, 103 samples were selected for modeling. The dataset is divided into training and test sets according to the ratio of 4∶1, and five algorithms, including Savitzky-Golay smoothing (SGS), Standard Normal Transform (SNV), Multivariate Scattering Correction (MSC), First-Order Derivative (FD), and De-Trending (DT), are applied to preprocess the spectral data and build the SSC full-band prediction model. The competitive adaptive reweighting algorithm (CARS), successive projection algorithm (SPA), and random frog hopping algorithm (RF) are used for feature band selection, and the SSC feature band model based on the partial least squares regression algorithm (PLSR) and support vector machine regression algorithm (SVR) is established. The results show that: SNV, MSC, and FD achieved better preprocessing results, and the prediction accuracy of the eigen-band model was significantly improved compared with the full-band modeling. The “SNV-CARS-PLSR” model built by SNV preprocessing combined with CARS feature extraction performs optimally, The training set coefficient of determination ($R_{\text{C}}^2$), training set root mean square error (RMSE_C), test set coefficient of determination ($R_{\text{P}}^2$), test set root mean square error (RMSE_P), and residual prediction deviation (RPD) were 0.869, 0.219, 0.858, 0.191, and 2.715, respectively. Compared to the SNV-preprocessed full-band model, the “SNV-CARS-PLSR” model improves the $R_{\text{P}}^2$ of the test set by 12.3%. Comparing the different feature band modeling methods, the “SNV-CARS-SVR” model based on SVR is slightly better than the “SNV-CARS-PLSR” model based on PLSR. The “SNV-CARS-SVR” model has an $R_{\text{P}}^2$ of 0.881, an RMSE_C of 0.207, an $R_{\text{P}}^2$ of 0.869, an RMSE_P of 0.185 and an RPD of 2.843. This study can provide technical support for rapidly detecting SSC in fresh corn cobs based on near-infrared spectroscopy.