diff --git a/python/pyspark/ml/image.py b/python/pyspark/ml/image.py
index 2d86c7f03860c..45c936645f2a8 100644
--- a/python/pyspark/ml/image.py
+++ b/python/pyspark/ml/image.py
@@ -40,6 +40,7 @@ class _ImageSchema(object):
     def __init__(self):
         self._imageSchema = None
         self._ocvTypes = None
+        self._columnSchema = None
         self._imageFields = None
         self._undefinedImageType = None
 
@@ -49,7 +50,7 @@ def imageSchema(self):
         Returns the image schema.
 
         :return: a :class:`StructType` with a single column of images
-               named "image" (nullable).
+               named "image" (nullable) and having the same type returned by :meth:`columnSchema`.
 
         .. versionadded:: 2.3.0
         """
@@ -75,6 +76,23 @@ def ocvTypes(self):
             self._ocvTypes = dict(ctx._jvm.org.apache.spark.ml.image.ImageSchema.javaOcvTypes())
         return self._ocvTypes
 
+    @property
+    def columnSchema(self):
+        """
+        Returns the schema for the image column.
+
+        :return: a :class:`StructType` for image column,
+            ``struct<origin:string, height:int, width:int, nChannels:int, mode:int, data:binary>``.
+
+        .. versionadded:: 2.4.0
+        """
+
+        if self._columnSchema is None:
+            ctx = SparkContext._active_spark_context
+            jschema = ctx._jvm.org.apache.spark.ml.image.ImageSchema.columnSchema()
+            self._columnSchema = _parse_datatype_json_string(jschema.json())
+        return self._columnSchema
+
     @property
     def imageFields(self):
         """
diff --git a/python/pyspark/ml/tests.py b/python/pyspark/ml/tests.py
index 1af2b91da900d..75d04785a0710 100755
--- a/python/pyspark/ml/tests.py
+++ b/python/pyspark/ml/tests.py
@@ -1852,6 +1852,7 @@ def test_read_images(self):
         self.assertEqual(len(array), first_row[1])
         self.assertEqual(ImageSchema.toImage(array, origin=first_row[0]), first_row)
         self.assertEqual(df.schema, ImageSchema.imageSchema)
+        self.assertEqual(df.schema["image"].dataType, ImageSchema.columnSchema)
         expected = {'CV_8UC3': 16, 'Undefined': -1, 'CV_8U': 0, 'CV_8UC1': 0, 'CV_8UC4': 24}
         self.assertEqual(ImageSchema.ocvTypes, expected)
         expected = ['origin', 'height', 'width', 'nChannels', 'mode', 'data']