Sorting and recycling trash is a big deal for saving resources and cleaning up the environment. Plastic waste is a key part of this process, and we're getting better at finding it by using more than one type of image.

For a long time, we've used only one type of image to spot plastic waste. Now, we're starting to use two kinds of images together: one that we see with our eyes (RGB) and one with special light that can detect things we can't see (hyperspectral images). But even with this new method, we're not using all the info we could. That's why two new datasets were created to help researchers figure out how to sort plastic waste better.

An important part of this is picking the best hyperspectral bands from the images. To do this, a new feature band selection algorithm was made using the Activation Weight function. This helps cut down on the work needed to get, send, and figure out the data.

Next, we have a tool called the Selective Feature Network (SFNet) which balances info from different images and stages. It works with a special block called the Correlation Swin Transformer Block to smash together info from different images. This teamwork makes spotting plastic waste even better.

Tests showed that this new way of picking bands and using the Correlation SF-Swin Transformer got the best results. It scored 97. 85% and 97. 37% in two different experiments. If you want to see how it's done, check out the code and models here and the data here.

This way of doing things is a big step forward, but there's always more to learn. Next time you throw away a plastic bottle, think about how tech like this might help get it sorted and recycled.

Improving Plastic Waste Sorting with High-Tech Multi-Image Analysis

Actions