Extracting richer representation with CNNs