GrapeRefineNet: Deep Learning Network for Enhanced Grape Disease Recognition with ResNet-34 and Refined Object Detection

Syed Sohail Ahmed^1*Marriam Nawaz²Jamal Nasir Abdullah Alotabi¹

• ¹Qassim University, Buraydah, Saudi Arabia
• ²University of Engineering and Technology, Taxila, Taxila, Pakistan

Grapes are an essential agrarian crop that contributes considerably to the worldwide economy and food industry. However, the cultivation of grape plants is highly open to various abnormalities that can can profound effect on their growth and quality. Therefore, prompt and effective recognition of these diseases is vital for reliable management and prevention. Traditional methods like manual inspection of plants are often labor-intensive and prone to inaccuracies. To overcome issues of manual approaches, various automated methods are devised, however, still numerous challenges persist as these approaches struggle with distorted samples, inaccurate localization of infected regions, and vulnerability to image postprocessing attacks. To address these challenges, we propose an improved RefineNet framework specifically designed for grape plant leaf disease recognition and give it the name GrapeRefineNet. The model comprises two modules: the Anchor Refinement (AR) unit and the Object Detection (OD) module. The AR unit is concerned with boosting the localization of infected regions by refining anchor boxes, while the OD module ensures precise classification of detected areas. To extract relevant visual details, we utilize ResNet-34 as the backbone network by utilizing its efficiency and strong feature representation capabilities. The work is tested employing grape samples from a standard dataset named the PlantVillage. Our approach surpasses base, and latest methods by achieving a high mean Average Precision (mAP) of 90.64% and demonstrating superior performance in precision, recall, and accuracy. We have validated its robustness via huge experimental evaluation and prove that the technique presents a reliable solution for real-world agricultural applications.