Design of a lightweight recognition network for adult locusts and grasshoppers based on deep learning

Grassland locusts and grasshoppers play a vital role in driving the dynamic changes of grassland ecosystems. In this study, we propose a lightweight deep learning-based network model for accurate identification of locust and grasshopper genera. Two image datasets were constructed, each containing 60 genera of locusts and grasshoppers. To improve recognition accuracy and computational efficiency while reducing floating-point operations (FLOPs) and the number of parameters, we introduced the channel-wise principal-component attention (CPCA) attention mechanism module and replaced part of the EfficientNet convolution modules with GhostConv, which incorporates the efficient channel attention (ECA) attention mechanism, thereby developing the CGENet model. During training, transfer learning and the Adam optimization algorithm were employed, significantly enhancing accuracy. This study makes precise control of locusts and grasshoppers feasible, thereby helping to reduce the damage they cause to agricultural production.