Identification of Butterfly Species in the Wild Based on YOLOv3 and Attention Mechanism
摘要
To solve the problems of fine granularity of butterfly classification, low recognition efficiency and poor accuracy of existing models in the field environment, aim at the automatic recognition of butterfly species in the field, an embedded channel attention MultSE1D recognition network was proposed to improve the backbone network of YOLOv3 model on the basis of self-built hybrid data set. The network used multiscale to extract high-dimensional features, so that the network had a variety of receptive fields, and paid more attention to the local subtle differences between many subclasses of butterflies and the surrounding environment. One dimensional convolution was used to replace the compressed excitation layer to avoid dimensionality reduction of channel features, to reduce the model parameters and to improve the operation efficiency of the model. According to the above method, the final mean average precision (mAP) of the model was 83.2%. The results showed that the improved recognition network could effectively improve the accuracy of the original model to extract the butterfly image features and the learning ability of the detail features, and could provide an effective solution to the problem of identifying the species of butterfly digital images in the wild.