Banana Bruise Detection

Role: Co-Author · Team: 3 · Timeline: Nov – Dec 2025
Tools: MATLAB, Image Processing Toolbox, Color Thresholder, k-nearest neighbors
Overview
I co-developed a MATLAB computer vision pipeline that looks at a photo of a banana and decides whether it's OK or bruised. It segments the peel into healthy and bruised pixels by color, then classifies the banana with a k-nearest-neighbors vote against labeled examples. It was our final project for ECET 369 at Purdue, applying vision methods to an automated inspection problem.
What I did
- Segmented the banana from the background with an HSV saturation threshold, then cleaned the mask
(
bwareaopen,imfill,bwareafilt) and cropped to its bounding box - Tuned color thresholds in MATLAB's Color Thresholder app to separate bruised peel from healthy peel
- Turned each banana into two features, its healthy-pixel and bruised-pixel counts, and classified it with a 3-nearest-neighbors vote
- Measured accuracy on held-out test images and on the training set with leave-one-out
Challenges
Our biggest problem came from the data itself: the shadows the bananas cast in our photos. Standard color thresholding picked up those shadows, and shaded areas of healthy peel, as bruised regions. Rather than relying on a single bruise threshold to make the call, we used a k-nearest-neighbors classifier that compares each banana's pixel counts against labeled good and bruised examples, so shadow that shows up in every photo has less effect on the result.
Results
On 10 photos (6 for training, 4 for testing):
- Test set: 3 of 4 correct (75%)
- Training set, leave-one-out: 3 of 6 correct (50%)
The classifier recognizes good bananas reliably but tends to miss bruised ones. The biggest next steps would be a bruise threshold that ignores shadows, using the bruised fraction of the peel instead of raw pixel counts so image scale doesn't matter, and a larger dataset.