Recognizing Pests in Field-Based Images by Combining Spatial and Channel Attention Mechanism

Large scale pest recognition is one of crucial components in pest management in outdoor conditions, which is much more difficult than common object recognition because of the variational image acquisition direction, location, pest size and complex image background. To overcome the challenges, this study proposes a CNN model by combining spatial attention mechanism and channel attention mechanism to realize accurate pest location and recognition in field images. The proposed model consists of two major parts. Firstly, the module Spatial Transformer Networks (STN) is incorporated into a Convolutional Neural Network (CNN) architecture to provide image cropping out and scale-normalization of the appropriate region, which can simplify the subsequent classification task. The second one is called Improved Split-Attention Networks that is used to enable feature-map attention across feature-map groups. The proposed model is evaluated on three different datasets: Li’s dataset (10 species), proposed dataset (58 species) and IP102 dataset (102 species), achieving the classification accuracies of 96.78%, 96.50% and 73.29%, respectively. Comparisons with five traditional CNN models and three attention-related state-of-the-art deep learning models show that the current method outperforms these previous models. Besides, to verify the robustness of this proposed model on different image resolutions, six datasets with different image resolutions are constructed and all accuracies exceed 92% with the image resolution of $400\times 267$ pixels reaching the optimal performance. All results show that the proposed method provides a reliable solution to recognize insect pest in field and support precision plant protection in agriculture production.