Hi any guys there can help me with the below problem?

I have to work on face recognition using SOM and MLP. The idea is to divide the face image into subblocks, use the subblocks to train a SOM, and find the clusters. This part no problem.

The problem comes in here: how do i use the SOM to train a MLP? My idea was to use all the best matching units of the subblocks, get their class label on the map by majority voting, then use the BMUs to train the MLP. Since the images are from 16 people, the no. of output nodes of the MLP will be 16. The no. of input nodes of the MLP is 4 since each subblock of the image is of dimension 4 (2 x 2). I repeat, all the subblocks are used to train the SOM.

After training of SOM and MLP, a test image is first divided into subblocks and then projected onto the SOM to find the BMUs. These BMUs of the test image will then be inputted into the MLP to be classified. Is this possible? If not how should i go about designinig the MLP? Can someone answer me first before i go into the details of the design? Many thanks.