Abstract: 【Aim】 Butterflies, as lepidopteran insects, are sensitive to their living environment and can be indicator species of the regional ecological environment. Automatic identification of butterfly species in the natural environment is of great significance to ecosystem stability. In the existing studies, there are few species and numbers of butterflies, and most of them take specimen images as recognition objects. Therefore, the butterfly image data set in the natural environment was constructed, and a butterfly species recognition model LDResNet based on the residual network was proposed in this study. 【Methods】 Firstly, deformable convolution was introduced to enhance the network feature extraction ability to different shapes of butterfly images, and to obtain more fine
-grained features. Secondly, the attention mechanism was embedded after the deformable convolution to increase the weight of butterfly features and reduce the interference of redundant information. Finally, the number of model parameters was reduced using an improved depthwise separable convolution. 【Results】 Experimenting on a self-built dataset, the LDResNet model achieved the average recognition accuracy of 87.61%, a 3.14% improvement over the original model, with only 1.04 MB of model parameters. 【Conclusion】 LDResNet has obvious advantages over other models in terms of average recognition accuracy and number of parameters, and this research model can provide technical support for automatic identification of butterfly species in natural environments.

Key words: Butterfly, residual network, deformable convolution, attention mechanism; depthwise separable convolution