diff --git a/README.md b/README.md
index abd5648..e22efe2 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 
 ## Automatically grade MCQ exams using optical mark recognition
 
-Autograder allows you to automatically grade MCQ exams. It is written in pure
+`autograder` allows you to automatically grade MCQ exams. It is written in pure
 rust and runs both in the command line and the modern web browsers using wasm.
 
 ## Installation
@@ -108,23 +108,27 @@ triggering an update in the view. You can always look into the developer
 console, which has a rather verbose output to what is happening in the background.
 
 Afterwards you can download a zip file containing a CSV file with all the
-results and conveniently named reports like this ![example report](assets/sample_report.png).
+results and conveniently named reports like this:
+
+![example report](assets/sample_report.png)
 
 These image reports were the main motivation to develop this software: using OMR
 will always be a bit error-prone, especially when students use the bubble sheet
 in ways that is not intended.
 
-A green circle means that autograder thinks that the selected bubble is the
+A green circle means that `autograder` thinks that the selected bubble is the
 circled one, which according to the key is correct. A red circle shows the
-correct answer, meaning that the selected bubble is elsewhere.
+correct answer, meaning that the selected bubble is elsewhere. An orange box
+indicates that `autograder` was not sure how to understand the choice and that
+manual grading is indicated.
 
 ### Using autograder from a mobile device
 
-If you only want to use autograder to grade a handful of bubble sheets, you can
+If you only want to use `autograder` to grade a handful of bubble sheets, you can
 do it like this:
 
 First, on a device with a large display, navigate to *Create Magic Link*. Here
-you can upload a key and a template and autograder generates a _very_ long link.
+you can upload a key and a template and `autograder` generates a _very_ long link.
 This link encodes all the template and key data and can be shared with anyone --
 most importantly yourself for usage on a mobile device. Bookmark that
 very long link with a descriptive name like "Stat101 Test 1" on your mobile device
diff --git a/assets/filled_out_example.png b/assets/filled_out_example.png
index 8b08d22..5871e34 100644
Binary files a/assets/filled_out_example.png and b/assets/filled_out_example.png differ
diff --git a/assets/sample_report.png b/assets/sample_report.png
index b956d32..549695a 100644
Binary files a/assets/sample_report.png and b/assets/sample_report.png differ
diff --git a/src/image_helpers.rs b/src/image_helpers.rs
index abb05a9..5ced9e7 100644
--- a/src/image_helpers.rs
+++ b/src/image_helpers.rs
@@ -80,6 +80,8 @@ pub fn fax_to_grayimage(data: &[u8], width: u32, height: u32) -> GrayImage {
         y += 1;
     });
 
+    // we don't trust binary images and erode them first
+    imageproc::morphology::erode_mut(&mut result, imageproc::distance_transform::Norm::L1, 1);
     result
 }
 pub fn binary_image_from_image(img: DynamicImage) -> GrayImage {
@@ -164,29 +166,23 @@ pub fn rgb_to_egui_color_image(image: &RgbImage) -> egui::ColorImage {
 }
 
 pub fn create_error_image(error_text: &str) -> GrayImage {
-    // Create a new 400x300 grayscale image
-    let mut image = GrayImage::new(800, 300);
+    let mut image = GrayImage::new(1200, 300);
 
-    // Fill with light gray background
     for pixel in image.pixels_mut() {
-        *pixel = image::Luma([240u8]);
+        *pixel = image::Luma([255u8]);
     }
 
-    // Load font from binary data embedded in the executable
     let font_data = crate::typst_helpers::BIOLINUM_BOLD;
     let font = ab_glyph::FontArc::try_from_slice(font_data).expect("Error loading font");
 
-    // Configure font scale (size)
     let scale = ab_glyph::PxScale::from(30.0);
 
-    // Calculate text position
-    let x = 20; // Padding from left
-    let y = 150; // Vertically centered
+    let x = 20;
+    let y = 150;
 
-    // Draw the error text
     imageproc::drawing::draw_text_mut(
         &mut image,
-        image::Luma([50u8]), // Dark gray text
+        image::Luma([0u8]),
         x,
         y,
         scale,
diff --git a/src/template_scan.rs b/src/template_scan.rs
index 1f389aa..69e2266 100644
--- a/src/template_scan.rs
+++ b/src/template_scan.rs
@@ -238,7 +238,21 @@ impl<'a> TemplateScan<'a> {
 
     pub fn set_transformation(&mut self) {
         let trafo = self.find_transformation();
-        self.transformation = trafo;
+
+        if trafo.is_some() {
+            self.transformation = trafo;
+            return;
+        }
+
+        // if this did not work, then in all cases known to us the image had too
+        // much white noise, so we erode it
+        imageproc::morphology::erode_mut(
+            &mut self.scan.image,
+            imageproc::distance_transform::Norm::L1,
+            1,
+        );
+
+        self.transformation = self.find_transformation();
     }
 
     pub fn find_transformation(&self) -> Option<Transformation> {