Merge pull request #7 from uavaustin/nas-filter

Added Method for Filtering Shapes/Not Shapes

target_finder/preprocessing.py

@@ -56,17 +56,16 @@ def find_blobs(image, min_width=20, max_width=100, limit=100, padding=20):
         else:
             has_mask = False
 
-        if (width < min_width or height < min_width or height >
-                max_width or width > max_width):
-            continue
-
         # align the contour with the cropped image
         rel_cnt = np.array(cnt)
         rel_cnt[:, :, 0] -= max(x - padding, 0)
         rel_cnt[:, :, 1] -= max(y - padding, 0)
 
         # Add the blob to the list without the image.
-        blobs.append(Blob(x, y, width, height, None, has_mask, rel_cnt, edges))
+        blob = Blob(x, y, width, height, None, has_mask, rel_cnt, edges)
+
+        if _is_shape_like_blob(blob, min_width, max_width):
+            blobs.append(blob)
 
     # Sort the contours with the largest area first.
     blobs.sort(key=lambda b: b.width * b.height, reverse=True)
@@ -82,6 +81,34 @@ def find_blobs(image, min_width=20, max_width=100, limit=100, padding=20):
     return blobs
 
 
+def _is_shape_like_blob(blob, min_width, max_width):
+    """Verify that this could be a blob containing a shape"""
+    width = blob.width
+    height = blob.height
+
+    # check provided restrictions
+    if min(width, height) < min_width or max(width, height) > max_width:
+        return False
+
+    # TODO: Uncomment once "Close contours for masking images" is complete
+    # must be closed cnt
+    # if not blob.has_mask:
+    #     return False
+
+    # check bbox ratio
+    size_ratio = max(width, height) / min(width, height)
+    if size_ratio > 2:
+        return False
+
+    # check solidity
+    hull = cv2.convexHull(blob.cnt)
+    solidity = cv2.contourArea(blob.cnt) / cv2.contourArea(hull)
+    if solidity < .65:
+        return False
+
+    return True
+
+
 def _crop_image(image, x, y, width, height, padding):
     """Crop the image, but not past the boundaries."""
     return image.crop((