Skip to content

jairolucas/Sisfrutos-Papaya

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Sisfrutos Papaya: a Dataset for Detection and Classification of Diseases in Papaya

In recent years, approaches based on machine learning, more specifically Deep Neural Networks (DNN), have gained prominence as a solution to computer vision problems in the most diverse areas. However, this type of approach requires a large number of samples of the problem to be treated, which often makes this type of approach difficult. In computer vision appli-cations aimed at fruit growing, this problem is even more noticeable, as the performance of computer vision approaches in this segment is still well be-low the performance achieved in other areas. One of the main reasons listed by the literature for the little evolution in this area is the lack of large data sets duly and manually annotated, which are mandatory for applications that use cutting-edge computer vision techniques such as DNNs. The present work aims to leverage research in this domain, creating a new dataset of im-ages, of an unparalleled size in the literature, with the main diseases and damages of papaya fruit (Carica Papaya). The proposed data set in this work consists of 15,179 RGB images duly and manually annotated with the posi-tion of the fruit and the disease/damage found within it. To validate our dataset, we use it to train a DNN-based classifier in the task of detecting diseases and defects in a papaya image. We recreated the old challenge “Man vs. Machine” comparing our classifier with a human ex-pert in a real environment. Our model reached an f1-score of 80.01%, while the overall performance obtained by the human expert was 67.3%

Important:

This database has been updated with new images and multiple disease tags on the same image. To download, please see the new work at: https://github.com/jhony2507/Sisfrutos-Papaya

Requisitos

  • Ubuntu 18
  • Cuda 10.1
  • Cudnn 8.0
  • OpenCV 4.5

Neural network model and examples for inference

DataSet

Inference

  • to make inferences : ./darknet detector test data/obj.data cfg/yolov4.cfg nameimage yolov4.weights

Sample images for inference testing

https://drive.google.com/drive/folders/1GhCxUPzlfXBJRIXsuwiDwkZY8aNduu1_?usp=sharing

Results

F1-score 0.801

Comments

The poor performance in detecting the Black Spot class surprised us, as it has very striking visual characteristics, which should favor the neural network. Upon further investigation, we detected that the poor performance is related to how, in some cases, the disease was noted in the ground truth. Black spot can occur in several disjoint regions of the fruit (several instances of the “Black Spot” class), but often the evaluator marks a large region of the fruit as being a single instance of that class. This makes the network detect multiple instances of this class outside the ground truth bound box. The images below exemplify this situation.

We are reviewing the notes of this class, and this review, as well as the results will be available on the project page.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published