"INDTomatoMSTD: A Multi-Scale Tomato Disease Dataset for Tiny and Large Lesion Detection Using Deep Transfer Learning"

"Early diagnosis of plant diseases is one of the critical variables in enhancing the crop productivity, food security and promoting sustainable agricultural activities. Tomatoes are most susceptible to various viral and superficial infections which appear in the form of tiny holes or abnormal discolorations spots on the surfaces of fruit. Due to differences in the size of the lesion, the light conditions, and natural difference in the texture of tomato fruits, it may be difficult to detect such symptoms. In response to such issues, a new dataset called INDTomatoMSTD (Indian Tomato Multi-Scale Tomato Surface Disease Dataset) has been created to aid the research on automated disease detection by means of the cutting-edge deep learning and transfer learning methodologies.The data set includes images of tomato fruit of high quality that was gathered at actual farms in the Jaipur area of Rajasthan, India. Four large vegetable producing regions, Sanganer, Amer, Chomu and Bagru, were sampled in terms of data collection. The pictures were taken in real-life farm settings to give it realistic variation in brightness and surrounding as well as the positioning of the fruit. The dataset is narrow in terms of surface-level symptoms of the viral infection and samples of tomatoes that are healthy, which allows the researchers to educate the models that can differentiate the latent disease patterns.There are two main classes in INDTomatoMSTD, namely, Healthy and Rejected. The rejected class includes the tomatoes with the viral infections leading to surface diseases. To improve the dataset and allow training a machine learning model with significant power, several sophisticated image processing and augmentation methods were used. Eleven preprocessing functions were used, such as rotation at several angles, Gaussian blur filters of various kernel sizes, Gaussian noise addition, and cropping, grayscale conversion, and other geometric manipulations. The specific aim of these techniques was to replicate the variations of imaging in the real world and allow multi-scale learning of disease patterns.The dataset used has been preprocessed and augmented, and now consists of 1274 processed images in the healthy group and 3780 processed images in the rejected (diseased) ones. All the pictures were reduced to 800 x 600 pixels to maintain consistency when training the model. Multi-scale aspect of the data set can be used to enable detection models to capture disease characteristics that are tiny lesions on the surface or larger patterns of infection, which is especially useful to current objects detection models and deep transfer learning designs like YOLO, EfficientNet, and Vision Transformer based models.The purpose of INDTomatoMSTD is to be useful in the development of intelligent plant health systems, precision agriculture technology, and automated quality inspection systems of vegetable supply chains. This data set can also be used to support the research aimed at localized crop disease detection models that can be applicable in Indian agricultural settings given that this data is a region-specific dataset that was collected under real agriculture."