Tea leaf disease detection using segment anything model and deep convolutional neural networks

• A custom CNN architecture was developed to generate highly informative feature maps that can aid in quick and accurate tea leaf disease identification. • The capability of zero-shot segmentation provided by Segment Anything Model (SAM) was explored to extract and isolate the diseased portions of the tea leaves. • An image segmentation based technique was developed and used to assess its impact while detecting diseased tea leaf region. • Three classifiers namely MLP, SVM and DT were tried and the optimal classification approach that works best with the extracted features was determined. Tea is an important beverage across many cultures. Diseases affecting tea leaves can adversely impact the integrity, production and cause substantial economic losses. Hence, detecting these diseases efficiently and accurately at an early stage is extremely crucial. The dataset used in this work consists of 6 categories to be trained, namely: Algal Spot, Brown Blight, Gray Blight, Healthy, Helopeltis and Red Spot. In our proposed method, a convolutional neural network is used in conjunction with advanced image preprocessing techniques for detecting and segmenting the infected tea leaf region. OpenCV was employed to extract the Region of Interest (ROI) and image cropping was performed to focus only on the leaf. In the process of cropping, the leaf was identified in the image, a bounding box was drawn around it and then it was finally cropped to maximize the leaf in the image. Further, the Segment Anything Model's (SAM) zero-shot segmentation capabilities were tested to segment and extract the diseased regions of the leaf. Also, the images were fed into a custom Convolutional Neural Network (CNN) model to extract the relevant features. These features were subsequently assigned to various classifiers like MLP, SVM, and Decision Tree classifiers to classify the diseases. The performance of each model was analyzed and compared. An accuracy of 95.06 % was achieved demonstrating that the proposed model has relatively higher accuracy in identifying the tea leaf diseases than many of the existing models.