Bakheet, Almudayfir, Farag: DLFL-TLDNet Reaches 95.3% Training Accuracy

Authors Samy Bakheet, Abdullah Almudayfir, and Aly Farag said that their hardware setup for model training utilized a workstation equipped with an NVIDIA GeForce RTX 3060 GPU featuring 12 GB of VRAM, an Intel Core i7-11700 CPU running at 2.5 GHz with eight cores, and 32 GB of RAM, all operating on Ubuntu 20.04 LTS.

Technical performance metrics for the Frontiers study showed a steady upward trend in both training and validation accuracy throughout the training process, eventually reaching 95.3% and 92.3%, respectively. The loss curves recorded during these experiments displayed a smooth, progressive decrease for both sets, reflecting stable convergence without oscillation or divergence. Furthermore, the absence of a widening gap between these loss curves suggests that the model avoided significant overfitting issues.

The way tomato crops are diagnosed has changed through several earlier studies. Alzubaidi et al. (2021) previously established a deep learning pipeline using the Village Tomato Dataset that achieved high accuracy across multiple disease classes. Subsequent research by Haq and Khattak (2021), Ghosal and Ghosh (2021), and Pérez-Ruiz and Romero (2021) confirmed that deep learning models can reliably distinguish between visually similar disease categories.

In 2022, researchers introduced a reliable framework that can both locate and classify tomato leaf diseases, achieving top results on benchmark datasets. More recently, Tang et al. (2023) developed PLPNet, a model that integrates attention mechanisms and multi-level feature extraction to improve detection robustness and generalization.

The broader ecosystem of tools supporting these advancements includes Bryan’s (2023) Tomato Leaf Disease Dataset, hosted on Roboflow Universe, which provides an open-source resource for training and benchmarking models under diverse imaging conditions. Developers have also utilized object detection frameworks like YOLOv4, introduced by Bochkovskiy et al.

(2020), and active learning strategies such as entropy-based sampling, proposed by Wu et al. (2022), to further refine disease identification capabilities. These methodologies address the reality that many high-performing convolutional neural networks remain computationally expensive, which limits their suitability for mobile deployment in agricultural settings where timely, localized analysis is required to prevent widespread crop damage.

Photo of author

Economy Desk

Economy Desk at Archyde.

Neandertal Gene Variant Linked to Smaller Occipital Lobe and Lower Schizophrenia Risk