In the traditional Guttman-Harris type image analysis, a transformation is applied to the data matrix such that each column of the transformed data matrix is the best least squares estimate of the corresponding column of the data matrix from the remaining columns. The model is scale free. However, it assumes (1) that the correlation matrix is basic and (2) that the data matrix is free of measurement errors. In this paper a more generalized model is developed that does not require these two assumptions. Computational procedures are suggested. (Author)