Deep learning and IoT-based framework for sesame plant identification and weed detection

Automated weed identification in agriculture is a prominent and promising technological intervention as it improves the overall crop productivity and yield. But finding weeds in dense sesame fields is still hard because of occlusion and changes in lighting. This work suggests a hybrid deep learning model that works with IoT to solve these problems. The method uses a Region-based Convolutional Neural Network (R-CNN) model to find sesame plants and weeds. Approximately 1,300 images with YOLO-formatted annotations were used for training and testing. This technique includes making region recommendations, preprocessing images, utilizing a CNN architecture to classify objects, training an SVM, and leveraging CNN features to improve item bounding boxes. The model's structure is made up of three parts: CNN fine-tuning, CNN + SVM, and bounding box regression. The process of preparing the data included finding and sorting regions of interest. Subsequently, these areas are sorted into groups based on how they overlap with real-world items. After that, the features that were found were used to make accurate bounding boxes around the items that were found, which led to the best classification. The model in question achieved 99.73% classification accuracy with a maximum bounding box IoU accuracy of 0.787 during the forecast procedure. The methodology delineated in this study possesses the potential for application in agricultural sectors for weed removal in farming fields.