ThresholdedReLU Orthogonal Layer Weight Regularized Densely Connected Convolutional Networks CNN for Strawberry Disease Prediction

Strawberry fruits will be affected by metabolic or climatic problems, causing a significant threat to the crop’s production and quality as well as raising the risk of strawberry illnesses. Traditional identification methods, however, also have high rate of assessment inaccuracies and operate inadequately in real-time. This is completely obvious that in today’s environment of increasing strawberry output and reliability requirements, the traditional methods for strawberry disease diagnosis, which primarily rely on individual perspective and untrained eye monitoring, cannot meet the expectations of people for strawberry disease identification and control. A more comprehensive system must be created in order to accurately identify strawberry fruit diseases and provide suitable treatment alternatives. With this intent, this paper recommends ThresholdedReLU Orthogonal Layer Weight Regularized DenseNet (TODN) CNN model to identify the damaged strawberry using the Fruit360 Disease Dataset from KAGGLE. The proposed TODN model uses Fruit360 Disease Dataset from KAGGLE with 2000 Strawberry fruit leaflets having 1000 damaged strawberry fruit and 1000 healthy strawberry fruit. The 3 X 3 convolutional layers, 3 X 3 softmax pooling layer, 3 X 3 Transition and Dense Layer and 1 X 1 Dropout and Classification Layer were used to create the TODN model. It was then trained using the learning rate of orthogonal regularised Gradient Descent and ThresholdedReLU activation layer. The TODN was designed using a DenseNet network and ThresholdedReLU activation with 0.5 dropout and weight decay of orthogonal regularization. The dataset for Strawberry disease was partitioned into training and testing data, and the training data were fitted to both the IAGLR model and other CNN models in order to compare the effectiveness of the proposed IAGLR model. Python was coded on a Gtx Geforce Tesla V100 Nvidia graphics card server with a batch size of 64 and 40 training epochs. Results of execution indicate that the suggested IAGLR model showcase the accuracy of 97.77%, Precision of 97.62%, Recall of 97.25% and FScore of 97.77% when compared with existing CNN models.