Hybrid Deep Learning Network for Automatic Diagnosis of Rice Leaf Diseases Using Multipath Refinement Network with Region Attention Mechanism-Based Segmentation Model

Across the world, rice is a widely cultivated crop, a source of energy, and a staple food for half of the world’s population. In general, rice serves as an excellent energy source based on its carbohydrate content that provides a significant amount of calories, and is the central component of daily meals. Also, rice contains high caloric food content, which provides adequate protein, less fat, and other minerals. Various factors, including soil quality, weather conditions, insects, viruses, pathogens, and temperature, affect rice quality. Earlier researchers have shown that plant disease has a big effect on rice crops, which leads to continuous financial loss in the agricultural sector. Pytopathologists are pursuing reliable and effective techniques for accurately diagnosing and managing disease-affected rice plants. Farmers utilize a classical disease recognition methodology that consumes more time for visual inspection, resulting in huge economic losses. The evolution in agricultural technology helps to detect pathogenic organisms in rice plants and also controls the disease significantly using deep learning models that show remarkable attention in the research of crop disease. However, deep learning techniques like CNN, RCNN, Mask-RCNN, and C-GAN models have emerged for detecting the disease in plants. Limits in data samples and the small size of affected regions in the deep learning models can gradually decrease the detection performance. Due to complex conditions and changes in resolution size, the consideration of existing techniques fails to provide accurate detection outcomes. So, this work uses the hybrid deep learning model to solve traditional problems. At first, the required images are gathered from the Paddy Doctor: Paddy Disease Classification dataset and given to the segmentation phase. The multipath refinement network with region attention mechanism (MRNet-RAM) is developed to perform the segmentation process. Further, the images are entered into the classification using the optimized hybrid network (OHyNet), where the hybrid model comprises a pyramidal dilated DenseNet with an efficient capsule network. Moreover, to improve the classification accuracy, significant parameters are optimized in the OHyNet using the fitness divisible value of the dung beetle optimizer (FDVDO). Finally, the suggested system attains effective leaf disease classification outcomes. Then, the efficiency of the recommended technique is compared and evaluated with an existing model to showcase its superiority over others. From the simulation outcomes, the performance examination of the designed framework shows 94.52%, 94.54%, and 65.7% regarding accuracy, precision, and recall, respectively. The advancement of the simulation outcome helps the farmers to detect rice leaf disease promptly.