Title : Synthetic imaging for whitefly counting with few annotated real data
Abstract:
Counting insects in plants has many applications, such as the development of new insecticides. Deep learning has allowed to develop robust automation of the insect counting in images of leaves by density map estimation.
However, deep learning requires many annotated data. The annotation for insect counting is very time-consuming as precise location of all the insects must be indicated. We present a method that drastically reduces the annotation effort for training an insect counting model by creating many synthetic images and their annotations from very few real annotated data. The method is validated for whitefly counting in eggplant leaves.
This whitefly counting model is trained with leaf tiles of 256x256 infested with whiteflies. These synthetic tiles are created by, first, extracting whiteflies from 16 real annotated images of a complete leaf using the known whitefly location and classical image processing techniques for segmentation. Then, the extracted segmented whiteflies are visually inspected to discard blurred or poorly segmented ones. Finally, 50000 synthetic tiles are generated by pasting the extracted whiteflies in random tiles cropped from 59 images of clean eggplant leaves.
To validate this method, whitefly counting models are trained: a) only with tiles extracted from the 16 annotated real images; b) only with the 50 000 synthetic tiles and c) with both the real and synthetic tiles. The models are tested with
50 real annotated images. The process is repeated 5 times for different train and test subsets and the “median
(interquartile range)” of the Mean Average Error (MAE) and the coefficient of determination (R2) are reported. The
inclusion of the synthetic tiles greatly improves the results as it provides variability to the training data.
|
|
MAE |
R2 |
|
Only real |
17.84 (10.08-25.76) |
0.06 (0.03-0.41) |
|
Only synthetic |
8.64 (5.42-10.28) |
0.78 (0.38-0.85) |
|
Real + synthetic |
3.78 (3.11-4.97) |
0.85 (0.68-0.95) |
Audience Takeaway:
- The boost that deep learning has given to machine vision in order to make possible to digitalize and automatize complicated and time-consuming tasks such as counting insects in plants to measure the pest level.
- The big obstacle that is for the development of robust deep learning models the difficult availability of manually annotated data to train them, as the manual annotation task is usually very time-consuming and requires expert knowledge in the field.
- The big positive influence that the generation of synthetic annotated data has to get robust deep learning models requiring few manually annotated real data.
- The demonstration of the previous statements for a real use case of counting whiteflies in images of eggplant leaves taken in the wild. The synthetic images allow to drastically reduce the effort needed to get a deep learning model that counts the whiteflies with very good performance.
