Embedded plant recognition: a benchmark for low footprint deep neural networks

Plant recognition is a challenging task due to the following elements: many classes, the variability of organs within a species, the similarity of organs between species, the shooting conditions, etc. There exist many mobile applications for plant recognition but most of them require an Internet connection to send the image to a server that will compute the recognition and, send the result back. However, in nature, in the mountains or in the forest, Internet connections are very often poor or non-existent. The only embedded plant recognition application is InterFolia based on SqueezeNet network but is it the best architecture to recognize plants? In this work, we propose to compare main existing networks that can be embedded allowing the recognition of plants from their organs (leaves, flowers/fruits, barks). The aim is to study how these networks behave in the face of this difficult problem to highlight their advantages and disadvantages in this context. The elements of comparison are not only the performance of the networks but also their memory impact and their inference time on computer and smartphone. Such elements could be extended to other applications in similar contexts, such as embedded phenotyping. We also propose a dataset with 477 plant classes that we make available to the scientific community
¹
.