Using Deep Learning to Accurately Identify Moss

DZone 's Guide to

Using Deep Learning to Accurately Identify Moss

Moss is the most interesting thing, right? Well, it's actually pretty awesome when you consider that it inspired new advancements in AI: using 4D cameras and image recognition to successfully identify moss.

· AI Zone ·
Free Resource

Machine vision has come on an awful lot in recent years, and I've covered a number of the innovations that have helped that march. For instance, only recently I covered the creation of a new 4D camera that can generate images across 138 degrees of information. The team believes that the devices will provide a huge boost to the capabilities of a range of autonomous technologies that need to understand the environment in which they operate.

Even then, though, some objects remain incredibly challenging to pick up, with things like grass something that has traditionally stumped the algorithms. That's because such objects are amorphous and therefore hard to define.

This is especially crucial when applied to things such as the analysis of trees or crops from aerial imagery — whether that's for farming or utility maintenance work.

Seeing the Green

A team from Kyoto University believe they have found an answer. In a recent paper, they describe a technique that utilizes deep learning to recognize amorphous plants. The researchers taught the algorithm how to identify various types of moss in the hope that it would help them identify various other amorphous forms too.

Three distinct forms were used, with photos taken of the mosses both in places where they dominate but also where they are present in generally non-mossy areas. The hope is that the algorithm will be able to not only identify the moss but also distinguish it from other objects in the images.

The algorithm was trained using 90,000 images, of which 80% were used to train the system and 20% were used to test it afterward. When the system was put through its paces, it was rather good at accurately spotting the moss. What's more, when shown images with multiple types of moss, it was able to identify each moss with an accuracy of around 90%.

Interestingly, it seemed to perform better for some mosses than others, with Polytrichum identified 99% of the time, whereas Hypnum proved much harder and was only correctly identified 74% of the time. The authors believe this is because Hypnum is more amorphous than the other types of moss, with less clearly defined features.

The team hopes next to further develop the algorithm by training it on pictures of moss taking at various times of the year. Likewise, the white balance on the camera used to capture the image could be standardized to ensure a more accurate capture of each moss.

Even so, the results are hugely promising and the team is confident that this could be applied to the analysis of aerial imagery to better identify various types of flora, which could have a range of valuable use cases.

The team plans to release the algorithm into the wild via an app that will help users identify various forms of moss from their smartphone, with the app primarily targeted at gardeners.

ai, deep learning, image recognition, machine vision

Published at DZone with permission of Adi Gaskell , DZone MVB. See the original article here.

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}