Teaching a camera to find an apple’s stem cavity

A sorting machine can recognize an apple and still be unsure which way it is facing. For orientation-aware handling, it may need to locate more than the fruit’s outline: the stem cavity, where the stem attaches, and the calyx, the flower end. These small features can be harder to identify than the fruit itself, especially when apples differ in color and surface texture.

In a manuscript submitted to Sensors for possible publication, Alican Güneş and colleagues ask which changes to an object-detection model help it find these landmarks. Their results point to one promising adjustment: replacing a component in the model with a lightweight attention module. But the tests do not establish how the approach would perform on a working sorting line.

A controlled test of model choices

The researchers collected and manually annotated 1,506 apple images, marking three classes: fruit body, stem cavity and calyx. Images were taken indoors at a fixed camera distance and under consistent lighting. The authors divided them into 1,054 training images, 301 for validation and 151 for testing.

They started with YOLO11s, a small version of the YOLO object-detection model, using its default training settings. Then they changed one factor at a time: image resolution, data augmentation (transformations applied to training images), optimizer settings and model architecture. They also tested different confidence thresholds, which set how sure the detector must be before reporting a landmark.

Changing one factor at a time helps compare individual choices, but this was not an exhaustive search. The test set contained 151 images from the same controlled collection. The authors say the dataset and implementation code are proprietary, which limits independent reproduction of the work.

Finding the small feature was harder

On the held-out test images, the baseline detector did best on the fruit body and less well on the smaller landmarks. Its reported F1 score—a measure that balances precision and recall—was 0.9901 for the fruit body, 0.9160 for the calyx and 0.7918 for the stem cavity. The stem cavity was the more difficult target in this evaluation.

Increasing image resolution, trying six augmentation variants and manually changing the optimizer did not improve the baseline’s overall F1 of 0.9047. Some changes improved particular class scores or other metrics, but none was a consistent overall improvement.

The strongest candidate replaced a YOLO11s backbone block called C2PSA with Coordinate Attention. In simple terms, this module helps the model weight visual information according to its position. The resulting model had an overall F1 of 0.9038, close to the baseline, and the study’s highest mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5): 0.935. Its stem-cavity F1 was 0.8217, the highest reported for that class in the study. It also had 10.2% fewer parameters than the baseline. That is not an across-the-board improvement: its stricter localization score, mAP@0.5:0.95, was lower than the baseline’s.

A confidence-threshold sweep identified 0.15 as the threshold with the highest test-set F1 in that sweep. Combining the attention change with stronger color augmentation did not add to its benefit. Instead, the combined model’s overall F1 was 0.8775 and its stem-cavity F1 was 0.7319—both below the attention-only model.

What the results do—and don’t—show

For this image collection and setup, the tested architectural change was more promising than the tested adjustments to resolution, augmentation or optimizer. This does not show that Coordinate Attention will improve all apple detectors, or that the selected threshold will be best with other cameras, lighting or sorting equipment.

The authors note important limits: one controlled imaging setup, a moderate-sized dataset, only one backbone family and a limited set of tested changes. They also report that their inference-speed measurements varied between sessions, so those measurements should not be treated as a reliable deployment benchmark.

There is a further caution about the threshold result. The paper selects 0.15 using the held-out test set. Using test data to choose a setting can make the final estimate less independent; a separate test set would provide a stronger check of performance after that choice. The authors also tested only one combination of changes, so other combinations could behave differently.

The manuscript is described as submitted to Sensors for possible publication. The supplied source does not establish its review status. For now, the work is best understood as a focused engineering experiment: one model change improved a difficult landmark score in a controlled image collection, while combining it with stronger color augmentation reduced performance. Evaluation on independent images and varied equipment would be needed to determine whether the approach transfers to real sorting conditions.