An improved deep convolutional neural network for detecting plant leaf diseases

This research proposes a novel dense convolutional neural network with five dense blocks (5DB-DenseConvNet) for the detection of plant leaf diseases. Five dense blocks and four transition layers were used to develop the 5DB-DenseConvNet. The 5DB-DenseConvNet was trained and tested using a plant leaf disease image dataset. Such advanced data augmentation techniques as color augmentations, geometric transformations, neural style transfer, principal component analysis, color augmentation, random erasing, and super-resolution generative adversarial network were used to increase the dataset size. The dataset contains 97 diseased and healthy image classes of 21 plants and comprises 145,500 images of healthy and diseased plant leaves. The Bayesian searching technique was used to optimize the hyperparameter values of 5DB-DenseConvNet. Compression factor was introduced in the transition layer to improve the compactness of the 5DB-DenseConvNet. The 5DB-DenseConvNet was trained on a graphics processing unit environment for 1000 epochs. The trained 5DB-DenseConvNet performed at an average classification accuracy of 99.36% and at an average precision of 98.83% on the test dataset. The experimental results demonstrate that the performance and reliability of the proposed 5DB-DenseConvNet on plant disease detection are superior to prior transfer learning techniques.