Features of pyramid dilation rate with residual connected convolution neural network for pest classification

Abstract

The recent survey shows at least 1 in 8 people suffer from either malnutrition or hunger. The world’s growing population is raising more concerns about increasing food productivity. Harmful pests destroy a large percentage of the food grains during the initial stage of grains growth. The qualitative productivity of food increases as measure to control increases. Also, crop yield significantly increases when complete pest management measures, such as crop rotation, biological treatments, and targeted pesticide application, are used in agriculture. The proposed methodology mainly focuses on identifying and classifying harmful and beneficial pests to increase food productivity. The correct identification of harmful pests aids in the early information to the farmers, thereby increasing grain productivity. This paper proposes a Features Pyramid Dilation Residual Convolution neural network (FDPRC net) for identifying and classifying pests in the tomato pest, Wang, Xie, and pest dataset. The three concrete deep learning approaches include without skip, dense skip, and residual skip connection architectures. FDPRC net consists of the feature of a dilation pyramid convolution block (FDPCB). The element-wise addition of feature representations from each FDPCB’s increases the representational capability of the FDPRC net. The dilated convolution blocks increase the field of view without increasing the parameters. The residual and dense skip connections have mainly reduced the vanishing gradient problem in the proposed deep neural network. Multiple training and test sets using various data sets have shown that the proposed FDPRC net performs better than the state-of-the-art techniques. The residual skip connection with FDPRC net achieved greater classification and identification accuracy of 98.12%, 97.43%, 93.98%, and 93.46% for the tomato pest, wang dataset, Xie dataset and pest dataset respectively.