Deep Multibranch Fusion Residual Network for Insect Pest Recognition

Earlier insect pest recognition is one of the critical factors for agricultural yield. Thus, an effective method to recognize the category of insect pests has become significant issues in the agricultural field. In this paper, we proposed a new residual block to learn multi-scale representation. In each block, it contains three branches: one is parameter-free, and the others contain several successive convolution layers. Moreover, we proposed a module and embedded it into the new residual block to recalibrate the channel-wise feature response and to model the relationship of the three branches. By stacking this kind of block, we constructed the Deep Multi-branch Fusion Residual Network (DMF-ResNet). For evaluating the model performance, we first test our model on CIFAR-10 and CIFAR-100 benchmark datasets. The experimental results show that DMF-ResNet outperforms the baseline models significantly. Then, we construct DMF-ResNet with different depths for high-resolution image classification tasks and apply it to recognize insect pests. We evaluate the model performance on the IP102 dataset, and the experimental results show that DMF-ResNet could achieve the best accuracy performance than the baseline models and other state-of-art methods. Based on these empirical experiments, we demonstrate the effectiveness of our approach.

© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.