摘要
近年来,人工智能技术在害虫识别领域得到广泛应用。目前深度网络害虫识别方法仍存在计算量大、对复杂背景下的害虫识别效果差等问题。为了解决计算量大的问题,本文提出了一种新型轻量网络——WraNet。该网络利用二维离散变换模块对图像进行特征混合,并学习图像的强先验知识,例如尺度不变性、平移不变性和边缘稀疏性。这使得单层二维离散小波变换层达到多层深度神经网络的效果,从而减少了计算量和模型参数的大小。本文还提出了一种新的算法——WraNet-m,该算法通过软投票集成了WraNet、ResNet50和FPN网络模型,以进一步提升识别效果。WraNet-m算法在IP102和D0害虫数据集上的准确率分别达到了72.44%和99.52%,证明了集成方法的有效性和鲁棒性。
In recent years, with the promotion of agricultural informatization, artificial intelligence techniques have been widely applied in the field of pest recognition. However, current deep neural net-work-based pest recognition methods still face challenges such as high computational complexity and poor performance in complex background scenarios. To address the issue of high computation-al complexity, we propose a novel network called WraNet. This network employs a two-dimensional discrete transform module for token mixing and learns strong prior knowledge of the image, such as scale-invariance, shift-invariance, and sparseness of edges. It is worth noting that we also propose a new algorithm, WraNet-m, which combines WraNet, ResNet50, and FPN models through soft voting for further performance improvement. The WraNet-m algorithm achieves accuracies of 72.44% on the IP102 pest dataset and 99.52% on the D0 pest dataset, approaching state-of-the-art results on both datasets, thus demonstrating the effectiveness and robustness of the ensemble method.
出处
《图像与信号处理》
2024年第1期33-46,共14页
Journal of Image and Signal Processing
