To investigate the feasibility of discriminating explosive type with RGB image of fireball of explosion, the curves of time vs area with intensive light on the images were obtained on the basis of a lot of photographs of explosion of condensed explosive with small dosage (30 g) by the high speed photography, and multi-dimension eigenvector space to describe firelight of explosion of condensed explosives was built by using feature parameters of configurations of curves of area vs time. And finally the linear three-type pattern recognition with optimal eigenvector was made by applying three-type Bayes′s decision rule and generalized fisher criterion. The test data of TNT, RDX, HMX were discriminated. Results show that the inaccuracy is 10.07% of pattern recognition built by the training set, while it is 17.48% of pattern recognition with using Jackknifing method in the total sample sets.