Convolution network model based leaf disease detection using augmentation techniques

Abstract

AbstractAgriculture plays a vital role in a country's economy. Thus, better yields of crops are required for growth of agriculture‐based industries. Deep Learning (DL) techniques have led to remarkable achievements in image classification and recognition. However, DL networks rely heavily on large data sets to prevent overfitting. Image augmentation is one of the DL techniques gaining attention in avoiding the risk of overfitting. The most common Image augmentation techniques like rotation, zoom, and shift used in the existing research allow to generate new images from the original set and increases the images quantity but cannot minimize the misclassification error. The present research can provide a better solution to provide sufficient quantity of training images to a convolutional neural network model to handle the overfitting and classification problems. Therefore, two learning algorithms image preprocessing and transformation algorithm (IPTA) and image masking and REC‐based hybrid segmentation algorithm (IMHSA) are proposed to address the problem of limited dataset and convolutional neural network model overfitting during classification. IPTA is an adaptive supervised learning approach to transform the original images into augmented ones and IMHSA is an unsupervised approach for Red, Green, Blue (RGB) image segmentation. Later, the Histogram threshold technique is applied to form all the possible regions used to split the diseased leaf into comparable regions. A novel convolutional neural network model is also proposed to evaluate the performance of the IPTA approach. The model is trained on two independent datasets, one generated before and one generated after IPTA was applied. Plots of precision and loss functions are used to assess the acquired results. The experimental results demonstrated that before using IPTA, the training accuracy was 83%, while the validation accuracy was 65%. After using IPTA, the proposed model attained a training accuracy of 74% and a validation accuracy of 73%, thereby solving the overfitting problem. The experimental results proved that the proposed model outperforms while classifying the RGB images with the support of image augmentation.