一种迁移学习和可变形卷积深度学习的蝴蝶检测算法

Aiming at the demand of butterfly multi-features recognition, and the problems of low precision and efficiency of butterfly detection in ecological environment, a butterfly detection with deformable convolution depth neural network based transfer learning is proposed (TDDNET). Firstly, the ResNet-101 convolutional layer is reconstructed by using the deformable convolutional model, which can reinforce the learning of feature extraction network for butterfly features. At the same time, this algorithm is combined with the region proposal network (RPN) to construct a two-classes detection network named DNET-base. Next, on the DNET-base to build TDDNET, the subnetwork RPN is used to guide the deformable sensitive position RoI pooling layer, which can obtain the scores feature map and the multi-scale object location. Then, we use the Soft-nms to obtain better detection results. Finally, the model after DNET-base training is transferred to the TDDNET, and fine-tuning the TDDNET multi-classification parameters. In testing datasets which have 854 images, the butterfly mAP_0.5 of the proposed algorithm is 0.9414, mAP_0.7 is 0.9235, the detection rate (DR) is 0.9082 and the classification accuracy (ACC) is 0.9370. The experiments demonstrate that the proposed algorithm outperforms the state-of-the-art model in the same hardware environment. The results show that the proposed algorithm can detect butterflies with high accuracy.