Using near-infrared spectroscopy technology combined with synergy interval partial least square (SiPLS), competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA) and uninformative variable elimination (UVE) feature extraction methods, the universal detection models of blueberry sugar content were established by deep belief network (DBN) to achieve on-line non-destructive rapid detection. The near-infrared spectra of 280 blueberry samples of “Bluecrop” and “Reka” were collected, and the sugar content of blueberries was determined by a hand-held refractometer. Firstly, using the outlier samples detection based on joint X-Y distances (ODXY) method to detect abnormal samples, there were 2 and 4 abnormal samples from Bluecrop and Reka respectively. After eliminating 6 abnormal samples, the remaining 274 samples were divided into a training set and test set in a ratio of 3∶1 by the sample set partitioning based on the joint X-Y algorithm (SPXY). Secondly, Compared and analyzed the improvement effect of Savitzky-Golay smoothing (S-G smoothing), centralization, multiplicative scatter correction and other pretreatment methods on the original spectrum of blueberry. Using SiPLS to reduce the spectral dimension and filter the characteristic band and using SPA, UVE and CARS, choose characteristic wavelengths again. We established partial least square regression (PLSR) and DBN models with the optimal characteristic wavelengths. The results showed that the optimal pretreatment method of the blueberry sugar content near-infrared detection model was S-G smoothing, the optimal band of blueberry sugar screening by SiPLS method were 593~765 and 1 458~1 630 nm, and the UVE algorithm was used to select 159 optimal wavelengths from 346 variables screened by SiPLS. When establishing the DBN model of blueberry sugar content, we analyzed the influence of different hidden layer numbers on the detection model, and the root means square error of cross-validation (RMSECV) as a fitness function, the particle swarm optimization (PSO) was used to optimize the number of neurons in each hidden layer between ［1, 100］. It is found that the RMSECV of the DBN model reached the minimum value of 0.397 7 when the hidden layer was 3 layers, and the hidden layer node number was 67-43-25. Whether in the full spectrum or modeling characteristic wavelengths, the near-infrared DBN models of blueberry sugar content were superior to the conventional PLSR method. In particular, the characteristic wavelengths selected by the UVE method can greatly reduce the modeling variables, and the model accuracy was high. The correlation coefficient (RP) and root mean square error (RMSEP) of the optimal PLSR model were 0.887 5 and 0.395 9, respectively. TheRP and RMSEP of the optimal DBN model were 0.954 2 and 0.310 5. The research shows that the detection model of blueberry sugar content based on the characteristic wavelength extracted by SiPLS-UVE combined with the deep belief network method can better complete the accurate online analysis of blueberry sugar content, and the method is expected to be applied to the internal quality detection of blueberries and other fruits and vegetables.