When using statistical approach in texture analysis for image classification, more problems are to be met. Particularly gray level co-occurrence matrices approach is applied in discriminating different textures in images results better accuracy but with the high computational cost. In addition a single window classification for each pixel compromises the classification accuracy rate. This paper concentrates on image classification implementation by using a statistical method in pixel by pixel with maximum likelihood estimates. The experiments have been tested with the color images and results shows that the improvement in classification rates achieved, when compared to a single window classification.