LAGRANGE CONDITIONAL GENERATIVE ADVERSARIAL NETWORK (LCGAN) BASED IMAGE DENOISING AND MULTI-SCALE DUAL ATTENTION INCEPTION V3 (MSDAIV3) FOR TOMATO LEAF DISEASES DETECTION

Abstract

Agriculture has been the primary source of income for the majority of people in India. The impact of plant illness ranges from mild manifestation to the destruction of complete plantations that severely affect the agricultural economy. Therefore, early detection and diagnosis of these diseases are essential. During diagnosis, noise can be easily generated during image acquisition, transmission, and processing, which makes it challenging to extract disease features and reduces diagnosis accuracy. In this paper, new identification model is introduced for tomato leaf disease. Firstly, a Lagrange Conditional Generative Adversarial Network (LCGAN) is introduced to reduce the image noise interference. LCGAN, noise removed image is estimated by an end-to-end trainable neural network. LCGAN includes of encoder and decoder architecture so that it can generate better detection results. LCGAN is used to decrease the difficulty of extracting tomato leaf disease features in the identification network. Secondly, three-segment linear conversion is utilized to increase image contrast. Thirdly, Inception-V3 is proposed to extract abundant disease features. In Inception-v3, a batch normalization (BN) layer is inserted as a regularizer between the auxiliary classifier and the fully connected layer. Finally, Multi-Scale Dual Attention Inception V3 (MSDAIV3) model is introduced to measure the inter-class similarity and intra-class variability identification of tomato leaf diseases. The proposed model has been trained and tested extensively on real time dataset with six classes (Bacterial spot, early blight, Leaf Mold, Septoria leaf spot, Leaf Curl Virus, and Healthy). Denoising results of proposed model is compared to existing methods using metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Feature SIMilarity Index for Color images (FSIMc). The proposed approach shows its superiority over the existing methods using metrics like precision, recall, f-measure, and accuracy.