Plant Disease Detection Leveraging Latent Space based Mixing Methods for Image Data Augmentation

Vaishali Anup Suryawanshi^1*, Tanuja Kiran Sarode¹ and Sahil Ajay Adivarekar²

¹Computer Engineering Department, Thadomal Shahani Engineering College, Mumbai, India.

²MS (Pursuing), Computer Science and Engineering, The Pennsylvania State University, USA.

Corresponding Author E-mail: vaishali.suryawanshi@thadomal.org

DOI : http://dx.doi.org/10.12944/CARJ.12.3.34

Abstract:

Plant Disease Detection (PDD) is a crucial task in the field of agriculture since it directly affects plant production and subsequently the economy, social structure, and political scenario of any country. It has become one of the most researched topics due to its relevance and challenges involved. One of the challenges that the researchers face is the limited set of data for various plant diseases. Collecting the data on the field is a laborious and expensive task and labelling the images requires expertise in the domain. This paper addresses this issue by developing an Image Data Augmentation (IDA) technique that can be applied on the existing image dataset to generate huge number of images. The technique employed here uses feature space obtained using Hadamard transform which is real, orthogonal and symmetric. This transform is simple to implement, and its computational complexity is very less. This article proposes two mixing methods based on the Hadamard Transform. To test the effectiveness of the proposed methods three Convolutional Neural Network (CNN) Architectures viz. VGG16, VGG19 and ResNET50 are used on Plant Village dataset. The results of the proposed IDA methods are compared with the traditional augmentation methods. Analysis of the results show that both the methods have shown significant improvement over the traditional augmentation techniques on all three architectures with performance on ResNET-50 model being the best compared to VGG16 and VGG19.

Keywords:

CNN; Hadamard Transform; Image Data Augmentation; ResNET50; VGG16; VGG19

Download this article as: 

[ HTML Full Text]