LW-PWDNet: a lightweight and cross-terrain adaptive framework for early pine wilt disease detection

Abstract

Pine wilt disease (PWD) poses a severe threat to forest ecosystems due to its high infectivity and destructive nature. Early identification of PWD-infected pines is critical to curbing disease spread and safeguarding forest resources. In order to timely detect and prevent the spread of PWD and meet the requirements of edge computing devices for real-time performance and computational efficiency, this paper proposes a lightweight model LW-PWDNet. The backbone network reconstructs HGNetV2 to achieve efficient feature extraction. It decomposes traditional convolutions into more lightweight feature generation and transformation operations, reducing computational cost while retaining discriminative power. The feature fusion layer reconstructs the path aggregation network based on RepBlock and multi-scale attention mechanism, capturing fine-grained details of small lesions, so as to better capture the detailed features of small targets. At the same time, this paper designs a lightweight D-Sample down-sampling module in the feature fusion layer to further improve the model's detection ability for multi-scale targets. Finally, this paper designs a lightweight prediction layer LightShiftHead for this model. By strengthening the local feature expression, the detection accuracy of PWD in small targets is further improved. A large number of experimental results show that LW-PWDNet maintains a high detection accuracy of mAP 89.7%, while achieving low computational complexity of 5.6 GFLOPs and only 1.9M parameters, as well as a high inference speed of 166 FPS when tested on an NVIDIA RTX 4070 GPU with a 13th Gen Intel(R) Core(TM) i7-13700KF processor, using PyTorch 2.0.1 and CUDA 12.6, based on Python 3.9. This model can provide an efficient and lightweight detection solution for PWD in resource-constrained scenarios such as unmanned aerial vehicle inspections.