Pest Detection and Identification Guided by Feature Maps

Effective pest detection and identification are of great significance for agricultural activities, and morden machine learning methods, especially the deep neural network, undoubtedly provide convenient and effective ways for this issue. Most of the mainstream methods for pest detection require annotation of manual bounding box. However, the public pest datasets usually lack the labeling information. In this work, we propose a method for detecting pests with weak supervision. We obtain the key regions of the pest targets by deeply mining the information in the feature maps of the convolutional neural networks (CNNs). First, using the class activation mapping technique to generate activation maps of the target class during forward propagation of the model, and then obtaining the key regions of the target based on the hotspots extracted from the activation maps. In addition, our proposed method can be deployed as a general strategy to other CNNs, and experimental results demonstrate that our method can steadily improve the classification accuracy of the model. By applying our method to the single Resnet50, we can achieve 74.27% classification accuracy on the publicly accessible IP102 dataset. Another attribution of this paper is that we propose a lightweight approach to find solutions for pest detection in agricultural automation applications.