C-scan images are formed by mapping the windowed A-scan energy levels into pixel values. This operation limits the resolution of the resulting image due to the blurring effect of squared summations and the mechanical positioning resolution. However, one feature that is found in all such images that they are usually scanned in a raster scanning fashion (say horizontally). This collects most of the information in the rows of the scan. This paper explores the possibilities of applying a decomposition of the image in the row subspace and to filter out the stochastic effects as well as approximate the image parameters within a mean squared error criterion (MSE). Image is modeled as auto regressive structure with emphasis on the row-wise decomposition and suitable subspace deconvolution is implemented for suitable orders. Utilizing an initial guess, system parameters are identified recursively. Each recursion would imply an increase in signal to noise ratio, as can be shown by the entropy values for the image, and hence a better identification as well. The results are presented with artificial defects.