Combustion flame plays an important role in predicting endpoint contents of carbon and steel temperature’s in a basic oxygen furnace (BOF). Since the results of principal component analysis (PCA) exist errors, it was difficult to extract independent colors from the converted images. Different kinds of color swatch were selected from original flame images. By calculating the colors’ within-class distance and class distance between before and after transformation, the colors’ clustering performance was found that it keep consistent. Clustering method was used to determine color’s window size after PCA transformation. Through colors’ window, the characteristic of pixel number of sequence images change with time were extracted. Image data from two furnace A and B, smelted same type of steel, was used to make experiment. Results show that the color features, extracted by clustering analysis, is able to much better reflect the physical and chemical processes in steel-making.