Landfills play a huge part in contributing to our environment’s deterioration. In 2018, it was recorded that 80% of landfill waste are recyclable, yet despite this, over 11 million tons of recyclable garbage make their way into landfills each year. By making an easy way for users to manage their waste properly, we can greatly reduce damage. GreenVision hopes to spread awareness about this topic and have created a unique, innovative solution utilizing image recognition technology that can facilitate classification convenience.
What it does
Our website image recognizes the garbage and categorizes them according to its type, replicating how it functions at recyclable collection centres. They will then be grouped into their respective "conveyor lanes" (type of garbage) with efficiency.
How we built it
Item Recognition: Utilising Pytorch, we have created a Convolutional Neural Network that achieves a whopping 95% accuracy which can classify objects into 6 categories, namely cardboard, glass, metal, paper, plastic, and general trash. (most common recyclable materials)
Challenges we ran into
At first, we attempted to use Keras to train a CNN but it only achieved a disappointing 75% accuracy. As we were not familiar with Pytorch, its implementation cost us a lot of time and we met numerous roadblocks on the way, but we finally achieved a much more satisfactory accuracy. Moreover, there was a bit of difficulty with loading our custom pre-trained model in Pytorch which was a hassle, as well as deploying our Django Website on Heroku in a hurry.
Accomplishments that we're proud of
What we learned
We understood much more about using neural networks in Keras and Pytorch. We got more familiar with Python libraries like NumPy and Matplotlib which will definitely be useful in the future!
What's next for GreenVision
We'd love to add the feature to recognize multiple objects at once and label each object's status in the image. To further widen our audience, we would also hope to target children to reinforce eco-friendly habits and environmental-related education. Speaking of technicalities, we hope to implement the function of real-time item recognition with Computer Vision. Using more machine learning algorithms, we want to gather more user-based data to predict more accurate results. We also want to incorporate an actual user interface for different stakeholders to use the feature to their advantage.