Noise Robustness of a Texture Classification Protocol for Natural Leaf Roughness Characterisation

In the context of leaf roughness study for precision spray- ing applications, this article deals with its characterisation by computer vision techniques. Texture analysis is a pri- mordial step for applications based on image analysis such as medical or agronomical imaging. The aim is to classify textures after extraction of discriminating features. How- ever, this problem remains complex in the case of natural leaves because of changes in lighting, scaling or orienta- tion. There we consider a family of invariants from the frequency domain called Generalized Fourier Descriptors whose dimensionality is proportional to the spatial resolu- tion of the images. These features used with a Support Vec- tor Machines classifier lead to good results in terms of clas- sification error rate when the dimensionality is small but it gives more errors when the dimensionality increases; we use there different kinds of dimensionality reduction tech- niques (linear or non-linear) whose aim is to keep most in- formation in a vector of small dimensionality. It implies losses of information even if small. This is not the only source of losses, another one is related to the noise present in the images due to acquisition conditions and sensor sen- sitivity. We propose here to demonstrate the robustness of our method of classification despite these losses of infor- mation.