In the present study, the near infrared spectrum of freshwater fish was used to detect the freshness on line, and the near infrared spectra on-line acquisition device was built to get the fish spectrum. In the process of spectrum acquisition, experiment samples move at a speed of 0.5 m·s-1, the near-infrared diffuse reflection spectrum (900~2 500 nm) could be got for the next analyzing, and SVM was used to build on-line detection model. Sample set partitioning based on joint X-Y distances algorithm (SPXY) was used to divide sample set, there were 111 samples in calibration set (57 fresh samples and 54 bad samples), and 37 samples in test set (19 fresh samples and 18 bad samples). Seven spectral preprocessing methods were utilized to preprocess the spectrum, and the influences of different methods were compared. Model results indicated that first derivative (FD) with autoscale was the best preprocessing method, the model recognition rate of calibration set was 97.96%, and the recognition rate of test set was 95.92%. In order to improve the modeling speed, it is necessary to optimize the spectra variables. Therefore genetic algorithm (GA), successive projection algorithm (SPA) and competitive adaptive reweighed sampling (CARS) were adopted to select characteristic variables respectively. Finally CARS was proved to be the optimal variable selection method, 10 characteristic wavelengths were selected to develop SVM model, recognition rate of calibration set reached 100%, and recognition rate of test set was 93.88%. The research provided technical reference for freshwater fish freshness online detection.