To detect bruised samples from intact kiwifruits and to reduce the loss caused by decay fruits and cross-infection, the near infrared diffused reflectance spectroscopy and an Extreme Learning Machine (ELM) were coupled to establish a model to discriminate collided, pressed and intact kiwifruits during 10-day storage at 2 ℃. The effect of the discriminant models using the feature variables based on Uninformative Variable Elimination (UVE) and the characteristic wavelength by Successive Projection Algorithm (SPA) combined with UVE on simplifying model and improving prediction performance was compared. The results show that the collided samples can be distinguished easier than pressed ones from intact kiwifruits. Bruised kiwifruits can be recognized easier with the expansion of storage period. UVE-SPA-ELM model has optimal discriminant performance with a discriminant rate of 92.4% for total prediction set samples. This detection technique has a high measurement precision and applicability, and can be used to identify bruised kiwifruits nondestructively and rapidly.