Create a high performance CNN model for recognizing disease in tomato plants
In this paper We trained a deep CNN model to identify ill tomato plants using a relatively small dataset of 1000 images labeled by experts. To overcome the limited data challenge, we utilize a pre-trained model and transfer its learned knowledge to our task. This method is referred to as Transfer Learning. We used a model that is based on an Inception(Szegedy et al. 2016) that was pre-trained on the ImageNet dataset (Deng et al. 2009). The InceptionV3 model was adapted and fine tuned to our image classification task. We compared the performances of our model to ResNet50-based model (He et al. 2016) on the dataset. We also conducted a user study to establish the human level performance and to compare it to our model.
Our model outperformed the other model and the user study results
The Problem
1. Recognize TYLCV disease in Tomato plants.
2. Binary image classification task.
3. Dataset - Tomato plant images ranked by sickness level
Our Method
Image classification model:
-
Deep CNN - InceptionV3
Transfer learning based
-
ImageNet dataset
-
Architecture modification & Fine tuning
Conclusions
Novel method > ResNet50 based model
-
McNemar's test: p-value = 0.003
​Novel method > Human level
-
McNemar’s test: p-value = 0.049
Tomatoes are a very popular food, consumed by many people in the western society. In 2018, world production of tomatoes was 182 million tonnes. Tomato plants might suffer from different diseases which can cause considerable production and economic losses in the agriculture sector. Thus, there is a need to identify these diseases and trying to prevent them.
Recent advances in computer technology have vastly improved image identification abilities. Through the recent advances in Machine Learning, the performance of systems aiming to detect or recognize an object in images have been widely improving. Most notable image identification methods are deep Convolutional Neural Network (CNN) models - which can help in the task of detecting unhealthy plants in images.
One of the most devastating viral diseases of cultivated tomato plants is Tomato Yellow Leaf Curl Virus (TYLCV). This disease decreases the yield of up to 100%. In many regions, TYLCV is the main limiting factor in tomato production. Plants that are infected display common characteristics such as upward curling of leaves and yellowing of young leaves (Moriones and Navas-Castillo 2000). These characteristics can be identified by experts and treated.
Our novel method outpreformed both ResNet50-based model and human level performance
