Inception Adaptive Gradient L2 Regularized Learning Rate CNN for Strawberry Leaf disease Detection
M. Shyamala Devi, S. Shanthana, Bachina Hemasri
Abstract
Biochemical or environmental factors will impact strawberry plants, posing a serious warning to the crop's output and efficiency that increases the likelihood of strawberry diseases. Traditional identification techniques, however, perform poorly in real-time and have a high percentage of judgement errors. It is clear that in the present day of increasing strawberries production and quality requirements, the traditional techniques for strawberry disease recognition, which primarily rely on individual perspective and untrained eye monitoring, cannot meet the expectations of people for strawberry disease identification. A broader framework needs to be established in order to precisely recognize strawberry diseases and provide suitable treatment alternatives. Accordingly, utilizing the Strawberry Disease Dataset from KAGGLE, this research suggests employing the Inception Adaptive Gradient L2 Regularized Learning (IAGLR) CNN model to recognize strawberry calcium shortage leaf disease. The proposed IAGLR model was developed using convolutional layers of 3 × 3-dimension, maximum pooling layer, dropout layer of 1 × 1 dimension, and dense layer for each 3 × 3 convolution. Then, it was trained using the adaptive gradient optimizer and L2 regularized gradient descent learning rates. An Inception network, an Adaptive Gradient optimizer with 0.5 dropout, and weight decay of Gradient Descent L2 regularisation were used to build the IAGLR. In order to evaluate the performance of the proposed IAGLR model, the dataset for Strawberry illness was divided into training and testing data. The training data was mapped to both the IAGLR model and other CNN models. A Gtx Geforce Tesla V100 Nvidia graphics card server was used to programme Python, with a batch size of 64 and 40 training epochs. When compared to current CNN models, implementation outcomes reveal that the suggested IAGLR model exhibits FScore of 97.77%, precision of 97.62%, accuracy of 97.77% and recall of 97.25%.