基于高阶残差和注意力机制的轻量型作物害虫识别

Lightweight Recognition of Crop Pests Based on High-order Residual and Attention Mechanism

精准识别作物害虫对作物进行适时地防护和治理具有重要意义.在面向自然环境时,由于作物害虫体积小、与环境颜色的差异性不大,同时又面临着各种噪声和复杂背景等因素的影响,目前与深度学习相关的作物害虫识别模型存在难以兼顾识别准确率和鲁棒性的要求,难以部署在计算资源有限和低性能的移动端等缺陷.因此选取ShuffleNetV2网络结构中模型参数最少的ShuffleNetV2 0.5×为基准网络,设计了一个基于高阶残差和注意力机制的轻量型作物害虫识别模型(HOR-Shuffle-CANet).其中,前期的高阶残差可以为后面的网络层提供丰富的害虫特征,有效提高模型的识别准确率;坐标注意力(coordinate attention, CA)机制能够进一步抑制背景噪声和对作物害虫重点信息的关注,有效增强模型的鲁棒性;带标签平滑正则化(label smoothing regularization, LSR)的双稳态逻辑损失函数可以解决训练含噪数据集时逻辑损失函数的两个缺点,使得模型对噪声的适应能力更强.试验结果表明,HOR-Shuffle-CANet模型在自然场景中10类常见作物害虫图像的测试数据集上达到了91.22%的识别准确率,较基准网络提升了3.54个百分点.在保持轻量化计算的基础上,其识别准确率也高于现有的经典卷积神经网络AlexNet、VGG-16、GoogLeNet、Xception、ResNet-34和轻量级网络模型MobileNetV3-Small、EfficientNet-B0等.该模型具有识别准确率高、鲁棒性强和抗干扰性能好等特点,能够很好地适应作物害虫识别的实际应用需求.

Accurate recognition of crop pests is essential for timely crop protection and treatment. However, crop pests in natural environment are small in size and have almost the same color as the environment. Moreover, crop pest images are affected by various factors such as noise and complex background. Therefore, it is difficult for existing crop pest recognition models related to deep learning to balance the requirements of recognition accuracy and robustness and be deployed on mobile devices with limited computational resources and low performance. In this study, ShuffleNetV2 0.5×,which has the fewest model parameters in the ShuffleNetV2 network structure, is selected as the benchmark network, and a lightweight crop pest recognition model based on high-order residual and attention mechanism(HOR-Shuffle-CANet) is designed. Specifically, the high-order residual in the early stage can provide rich pest features for the subsequent network layer, which significantly improves the recognition accuracy of the model. The coordinate attention(CA) mechanism can further suppress the background noise and focus on the key information about crop pests, which effectively enhances the robustness of the model. The bi-tempered logistic loss function with label smoothing regularization(LSR) can solve two shortcomings of logistic loss functions in training noisy data sets and make the model more adaptable to noise. The experimental results show that the HOR-Shuffle-CANet model achieves a recognition accuracy of 91.22% on the test dataset of ten types of common crop pest images in natural scenarios, which is 3.54 percentage points higher than the benchmark network. On the basis of maintaining lightweight computing, its recognition accuracy is also higher than that of the existing classical convolutional neural networks such as AlexNet, VGG-16, GoogLeNet, Xception, and ResNet-34,as well as lightweight network models such as MobileNetV3-Small, EfficientNet-B0, etc. Due to its high recognition accuracy, strong robustness, and excellent anti-interference performance, the proposed model can meet the practical application requirements of crop pest recognition.