Optimizing Citrus Leaf Disease Detection: An Efficient Custom CNN Leveraging Efficient Training Parameters

Citrus leaf diseases threaten citrus estate profits. Early detection and proper diagnosis are essential for sickness recovery. Recent decades have seen advances in citrus leaf disease identification and categorization using deep learning approaches in image processing. We suggest employing pre-trained convolutional neural networks (CNNs) such as ResNet-50, VGG16, MobileNet-V2, InceptionV3, InceptionResNet-V2, DenseNet-201, and DenseNet-121 to automate citrus leaf disease diagnosis. A thorough data set of recognized citrus leaf images will be collected and pre-processed for this purpose. The datasets will include healthy and sick citrus canker, greening, and black spot leaves. Valuable features from leaf images is extracted utilizing deep learning models. The aforementioned models are effective and often used for image categorization. An efficient CNN model is proposed which trained on 596 images initially and after augmentation on 2800 images that are split into 3 categories: training, validation and testing. The data was split in as: 70% for training, 15% for Validation and 15% for testing purposes. This model will learn from some public data. Next, we compare results with prebuilt models. Finally, received 96% F1-score and 97.84% validation accuracy. Lastly, with a fair accuracy the proposed model has less training parameters than all other pretrained models which makes our model faster to classify the diseases.