leaf-disease-recognition

This project aims to study necrosis and pycnidia on a leaf. It is led by a PHD student in biology. My goal is to help her in the processing of her images. Following this analysis, we generate a dataset. This dataset will also be the subject of an in-depth study. This project is therefore the first step of her thesis.

Definitions :

• Necrosis : Death of tissue through injury or disease, especially in a localized area. Such an area is usually brown or black.

• Pycnidia : A pycnidia is a type of asexual reproductive structure found in fungi of the order Sphaeropsidales (class Coelomycetes) and lichens whose fungal component belongs to this order. The pycnidia is a spore-like concept of certain imperfect fungi (ascomycetes), usually globose or obpiriform in appearance (in the shape of a bottle or an inverted pear). Inside, very small asexual spores are formed, called conidia or pycnidias.

The dataset is composed about 1600 images and each leaf is composed of 4 portions of leaves.

Original image :

Here is how we proceed our analysis :

• We determine all the leaves on the image which is necessary for our result.csv file.
• We analyse all cropped leaves. On each leaf, we detect each necrosis area, which are necrotic areas of the leaf. Depending on the leaf, each necrosis may have a different color. To solve this problem, we use different masks. (See Analysis)
• Then, we detect the pycnidias. The areas calculated previously are very useful, because they allow to check if a pycnidia is coherent: if a pycnidia belongs to a necrotic area then it is a true pycnidia. (See Analysis)
• Finally, we create result.csv which is composed :
  • Columns from csv from PHD (columns containing researcher input data)
  • Name,
  • Leaf area in px and cm,
  • Number of necrosis areas,
  • Total area of necrosis areas in px and cm,
  • Number of pycnidia,
  • Total area of pycnidia areas in px and cm.

Analysis

Function : get_image_informations(directory, img, file_name, dpi, save)

Arguments :

• directory: main directory of analysis.
• img: TIF file to analyze.
• dpi: Number of pixels created on a one-inch area.
• save: save all images (cropped and analysed). True or False.

Necrosis treatment

Library used: OpenCV.

Different masks are used to determine all the necroses on a leaf.

1. Green necrosis
2. Green/Gray necrosis
3. Yellow necrosis

Then, the final mask is the assembly of these:

mask_merged = mask_yellow_necrosis + mask_green_necrosis + mask_gray_necrosis

Example :

Original

Green mask

Green/Gray mask

Yellow mask

Final mask

After having assembled all these masks, we apply these rules :

for necrosis in leaf:
  if necrosis area > 1000px:
      ratio = round(perimeter/area, 3)
      # Removing necroses with a marginal shape 
      # such as very elongated necroses
      if ratio < 0.25:
        # drawing necrosis
        i+=1
        necrosis_nb += 1
        necrosis_area += area

Pycnidias

Library used: SciPy.

One of our problematic is to detect leaf pycnidias which are technically small black dots. However, colors of small black spores are differents according to leaf background. We can't solve this just by color.

The common point between these spors is the shape. They can be assimilated to small circles.

That's why, we use convolution kernel.

Example :

After detecting the suspected pycnidia, we sort them according to some rules :

for pycnidias in leaf:
  if pycnidia is in necresis area: 
    if pycnidia color belong to authorized colors:
      # drawing pycnidia
      pycnidia_area += area
      pycnidia_number += 1