BEIJING, July 25, 2014 /PRNewswire/ -- As originally reported in Ironhands' blog, the consumer market has been looking for an app that can provide a precise image recognition function for years, and now there seems to be many apps dedicated to recognizing specific categories, including dogs, cats and flowers. And recently, a video clip posted on YouTube caught the customer's eyes: https://www.youtube.com/watch?v=Su6rY82_Z6M&feature=youtu.be

This video clip shows a fascinating competition between a well-trained botanist and a smartphone app, developed by Baidu, a Chinese search service provider, in order to determine if this dainty mobile application can beat a human expert in recognizing different types of flowers. The botanist, who has a doctorate from the Chinese Academy of Sciences, an academic institute in China, managed a 90 percent accuracy and, surprisingly, in the attempt to distinguish nearly 300 difference types of ornamental plants in 30 minutes, the app named "Baidu Photo Reader" managed an 82 percent recognition rate, which is quite impressive for an app. This app can now be downloaded from app stores like Google Play. 

Up till now, Baidu Photo Reader has the highest accuracy among all the similar apps, it even beat Google in a comparable test, which saw a Google-developed app achieving a recognition rate of 74.8 percent. Our test engineers then dug a little deeper into the background of this Baidu labeled app, and the information we got intrigued us: Baidu Photo Reader utilized the most massive and authoritative database and applied the technology of Deep Learning. When it comes to supervised learning, plants have proven to be a good category of training data.

There are also plenty of apps that provide an image recognition service in the area of botany; apps like LeafSnap UK, Plantnet, Audubon Trees, Geotress, FSC Hedges, and FSC trees. Tree key is fairly popular as well. Many of them use various verification methods. Audubon Trees and FSC trees, developed by Columbia University, identify species by recognizing their leaves. LeafSnap and Plantnet, on the other hand, can work with leaves as well as flowers, and are supported by French science department. 

Leafsnap UK, developed by Calouste Gulbenkian Foundation, can help users verify trees by leaves, flowers, and even fruits. Launched in 2011, this app now has a database of 23,915 high-definition pictures alongside with 5,192 reference photos taken by mobile devices. Pictures uploaded by users will be combined with the existing database. After two generations of entries, the app can now recognize tons of European vegetation, including 156 types of trees.

French Reseach Organizations (Cirad, INRA, Inria, and IRD) and The Tela Botanica Network also launched their image recognition app called Plantnet. It can provide an accurate identification for more than 3,700 kinds of plants around France, said its developer. Unfortunately, we haven't received any statistical data about its accuracy.

The truth is, the user experience of Leafsnap is rather poor. To correctly use the app, you have to put the plant's leaf on a white piece of paper. In contrast, users of Plantnet and Baidu Photo Reader can simply take a picture and recognize it. It is a pity that Plantnet cannot provide a static figure for comparison.

Image recognition requires both a dependable ability in acquiring large quantities of image data, and substantial technology accumulation in Machine Learning. The application of image recognition technology in vertical domains like ornamental plants is truly a reflection of the technical merit of companies. After a brief comparison of those apps above, it seems to be pretty clear that China may have outpaced the West in the field of image recognition.

Read the original post here: http://blog.sina.com.cn/s/blog_659f59780102uy3i.html

Source: Ironhands' Blog