Identification of linear models from quantized data: a midpoint-projection approach