diff --git a/ocrs/src/detection.rs b/ocrs/src/detection.rs
index 9337e94..1941ebf 100644
--- a/ocrs/src/detection.rs
+++ b/ocrs/src/detection.rs
@@ -5,6 +5,7 @@ use rten_tensor::prelude::*;
 use rten_tensor::{NdTensor, NdTensorView, Tensor};
 
 use crate::preprocess::BLACK_VALUE;
+use crate::tensor_util::IntoCow;
 
 /// Parameters that control post-processing of text detection model outputs.
 #[derive(Clone, Debug, PartialEq)]
@@ -165,22 +166,28 @@ impl TextDetector {
         // inputs, within some limits.
         let pad_bottom = (in_height as i32 - img_height as i32).max(0);
         let pad_right = (in_width as i32 - img_width as i32).max(0);
-        let grey_img = if pad_bottom > 0 || pad_right > 0 {
-            let pads = &[0, 0, 0, 0, 0, 0, pad_bottom, pad_right];
-            image.pad(pads.into(), BLACK_VALUE)?
-        } else {
-            image.as_dyn().to_tensor()
-        };
+        let image = (pad_bottom > 0 || pad_right > 0)
+            .then(|| {
+                let pads = &[0, 0, 0, 0, 0, 0, pad_bottom, pad_right];
+                image.pad(pads.into(), BLACK_VALUE)
+            })
+            .transpose()?
+            .map(|t| t.into_cow())
+            .unwrap_or(image.into_dyn().into_cow());
 
         // Resize images to the text detection model's input size.
-        let resized_grey_img = grey_img.resize_image([in_height, in_width])?;
+        let image = (image.size(2) != in_height || image.size(3) != in_width)
+            .then(|| image.resize_image([in_height, in_width]))
+            .transpose()?
+            .map(|t| t.into_cow())
+            .unwrap_or(image);
 
         // Run text detection model to compute a probability mask indicating whether
         // each pixel is part of a text word or not.
         let text_mask: Tensor<f32> = self
             .model
             .run_one(
-                (&resized_grey_img).into(),
+                image.view().into(),
                 if debug {
                     Some(RunOptions {
                         timing: true,
diff --git a/ocrs/src/lib.rs b/ocrs/src/lib.rs
index 4640faa..25a5a1c 100644
--- a/ocrs/src/lib.rs
+++ b/ocrs/src/lib.rs
@@ -11,6 +11,8 @@ mod log;
 mod preprocess;
 mod recognition;
 
+mod tensor_util;
+
 #[cfg(test)]
 mod test_util;
 
diff --git a/ocrs/src/tensor_util.rs b/ocrs/src/tensor_util.rs
new file mode 100644
index 0000000..7934a3d
--- /dev/null
+++ b/ocrs/src/tensor_util.rs
@@ -0,0 +1,37 @@
+use std::borrow::Cow;
+
+use rten_tensor::prelude::*;
+use rten_tensor::{MutLayout, TensorBase};
+
+/// Convert an owned tensor or view into one which uses a [Cow] for storage.
+///
+/// This is useful for code that wants to conditionally copy a tensor, as this
+/// trait can be used to convert either an owned copy or view to the same type.
+pub trait IntoCow {
+    type Cow;
+
+    fn into_cow(self) -> Self::Cow;
+}
+
+impl<'a, T, L: MutLayout> IntoCow for TensorBase<T, &'a [T], L>
+where
+    [T]: ToOwned,
+{
+    type Cow = TensorBase<T, Cow<'a, [T]>, L>;
+
+    fn into_cow(self) -> Self::Cow {
+        TensorBase::from_data(self.shape(), Cow::Borrowed(self.non_contiguous_data()))
+    }
+}
+
+impl<T: Clone + 'static, L: MutLayout> IntoCow for TensorBase<T, Vec<T>, L>
+where
+    [T]: ToOwned<Owned = Vec<T>>,
+{
+    type Cow = TensorBase<T, Cow<'static, [T]>, L>;
+
+    fn into_cow(self) -> Self::Cow {
+        let layout = self.layout().clone();
+        TensorBase::from_data(layout.shape(), Cow::Owned(self.into_data()))
+    }
+}