This classification model aims to identify items and categorize them based on how they should be disposed of. Using YOLOv11, this model fine-tunes previously trained datasets from Roboflow to fit new classes: recycle, trash, compost, and specialized disposal. This model is intented to be used to help people correctly dispose of their items and can be used for smart bins, which detected the item a person is holding and opens to the appropriate bin or for apps where the user can take a photo of the item and identify where it goes and how to dispose of it.

After merging the two datasets, I reorganized the classes into their new perspective classes seen in the table below.

This is the final class distributions

For the compost class, some images included items that could not be composted (ex, red meat). I reviewed all the images and moved non-compostable food waste to the trash category.

For the classes I chose to delete, the first one being plastic bottle caps, from Google searches, it is generally recommended to keep your bottle caps on your bottles when recycling, but for recycling plastic bottle caps, there were specific requirements on what size can and cannot be recycled. For example, according to the Seattle Public Utilities, loose bottle caps less than 3 inches in diameter go into the trash (Seattle Public Utilities). However, from the images alone, it’s difficult to interpret the size of the caps, so for less confusion in training, I choose to opt out of including them. For glass cups, they cannot be recycled and generally recommended to donate them; however, since there are only 3 images in this class, rather than adding a new “donate” class and risk significant class imbalance, I choose to delete them. Lastly, for both unlabelled litter and other plastic, it was difficult to identify these items, so I chose to delete them to minimize confusion.

For the recycle class, it came up to a total of 6,640 images, but because the rest of the classes were within the 1,000 range, and I wanted to try to prevent any false negatives and accuracy issues from imbalanced classes, I chose to delete 5,033 images from the class and ended up with 1,607.

Confusion Matrix

Train/Loss and Val/Loss Curves

From this model, the overall performance indicate high accuracy, with the top 1 accuracy being 0.962 and each class having a F1-score in the 0.9 range. From the confusion matrix, we can see a perfect diagonal, which indicates the model was able to accurately predict the items correctly. However we do see a few false negatives and positives, specifically where the model mixed up trash and specalized disposal is the highest, but this is very minor. Looking at the train/loss curve shows that the model is learning effectively, with the downward shape curve and decreasing spikes as the longer the model trains. From the metrics/accuracy_top1 curve we see large spikes towards the beginning of the training process but the spikes slowly decreases as it trains longer, which indicates the model is improving it’s accuracy and showing stable performance. For the validation loss curve, we see extremely high spikes towards the beginning, but the spikes slowly decreases as more training time is applied. Lastly, for the metrics/accuracy_top5, this shows up as a horizontal line at 1 due to the fact that I only have 4 classes. Overall, my model indicates high accuracy and no cases of overfitting, however, the model could benefit from longer training time to allow the curves to smoothen out further and reach an eventual straight horizontal line.