Visual regenerative fusion network for pest recognition

Automatic pest detection from agricultural crops is a challenging agricultural application. It has been observed that many farmers lose crop yields mainly due to pests that spread rapidly. An intelligent pest recognition system (IPRS) is an innovative application that assists farmers in increasing agricultural productivity by accurately detecting pests that may cause crop damage. It allows one to identify the type of pest and apply the appropriate pesticides in a timely manner. There is a large deviation in deep neural network (DNN) performance on clean images and naturally degraded images. This article proposes a novel data augmentation techniques and feature fusion techniques to improve the accuracy as well as the desired robustness. The proposed visual regenerative fusion network (VRFNet) fuses the multiscale features from global feature extraction (GFE) network and the visual regeneration (VR) network to capture the semantic details for identifying pests from the distorted images which improves the accuracy. The proposed patch-based augmentation approach can effectively simulate the environment in which the leaves can partially conceal the insects and helps in improving the robustness. The proposed model achieves an accuracy of 99.12 and 68.34 on the publicly available D0 and IP102 datasets, respectively. These results clearly demonstrate that the proposed fusion method achieved a high success rate for distorted and occluded images captured from the agriculture farms.