We are currently trying to detect disease in Mango leaves. We have collected the dataset by ourselves. Furthermore, we have tried VGG_16,19, ResNet, AlexNet, but we have achieved maximum accuracy of 73%. Tried pre-trained models. We increased the number of images using data augmentation( flipping, rotating, zooming etc.). We have a total of 1100 images only. 100 images for each class. Does CNN require more DataSet? Which CNN's we should use that work good on small datasets? I have attached the sample images. We tried segmented images as well. Do these segmented images have effect on the results? Should we work on better segmentation techniques? I am a newbie kindly be kind.🙃
Dear Ahmed Waheed,
Increasing the number of images in the dataset will help the model for a better learning process. However, the crucial part starts with the way of designing your model. Since a model cannot work perfectly for all the datasets, you need to design your system by optimizing its parameters for a specific task of yours. Moreover, to analyze your system performance, I recommend you check other performance metrics, and check the confusion matrix.
For the segmentation part, you could check the following paper in which we extracted several features from the segmented images to feed our model in order to improve the classification performance of our model. According to our results, this approach increased the performance of the classification.
Paper1:
https://www.researchgate.net/publication/343937860_A_Large-Scale_Dataset_for_Fish_Segmentation_and_Classification
Paper2: https://www.researchgate.net/publication/335368522_A_Comparative_Analysis_on_Fruit_Freshness_Classification
Best regards,
Dear Ahmed Waheed ,
You need to change your preprocessing technique, if you want to better accuracy you can segment the disease-affected portion in the preprocessing. You can apply DenseNet-201, M r101 X101*1, Inception ResNet-V3 algorithm, I think this performs very well in your dataset.
You can check this article for better understanding, I think it will help you a lot.
1. Article An Automated Convolutional Neural Network Based Approach for...
2. Article Recognizing Human Emotions from Eyes and Surrounding Feature...
3. Conference Paper Deep Learning Based Classification of Papaya Disease Recognition
As your dataset size is very small, my suggestion is not to use pretrained model if you want better test accuracy, I'm guessing your training accuracy is good and the model is overfitting. Otherwise, go for parameter tuning. May introduce some level of dropout also. By the way, could you tell me what type of segmentation method you've used?
- Badges
- Science topic