基于MobileViT-PC-ASPP和迁移学习的果树害虫识别方法

Fruit Tree Pest Identification Method Based on MobileViT-PC-ASPP and Transfer Learning

摘要:

为提高果树害虫识别效果，及时做好防治措施，本研究以6种对果树危害程度较大的害虫为研究对象，针对自然环境下果树害虫识别背景复杂、害虫目标小检测难度大、与不同类别间特征相似度高等问题，提出一种改进的轻量化MobileViT-PC-ASPP识别模型。该模型用PConv（Partial convolution）模块代替原模型MobileViT模块中部分标准卷积模块，其次修改MobileViT模块的特征融合策略，将输入特征、局部表达特征、全局表达特征进行拼接融合；删除网络第10层MV2模块和第11层MobileViT模块，使用改进空洞空间池化金字塔（Atrous spatial pyramid pooling，ASPP）模块进行代替，形成多尺度融合特征；此外，模型用SiLU激活函数代替ReLU6激活函数进行计算，最后基于ImageNet数据集进行迁移学习。实验结果表明，6类果树害虫图像识别准确率达93.77%，参数量为8.40×10.5，与改进前相比，识别准确率提高7.5个百分点，参数量降低33.86%；与常用害虫CNN识别模型AlexNet、ResNet50、MobileNetV2、ShuffleNetV2相比识别准确率分别提高8.25、4.78、7.27、7.41个百分点，参数量分别减少6.03×10.7、2.48×107、2.66×106、5.30×105；与Transformer识别模型ViT、Swin Transfomer相比识别准确率分别提高19.03、9.8个百分点，参数量分别减少8.56×107、2.75×107。本研究适合部署在移动终端等有限资源环境，并且有助于实现对复杂背景下小目标果树害虫进行识别检测。

Abstract:

In order to enhance the effectiveness of identifying pests in fruit trees and promptly implement preventive measures, focusing on six major pests that pose a significant threat to fruit trees, an improved lightweight MobileViT recognition model was proposed for the problems of complex background of fruit tree pest recognition in the natural environment, high difficulty of detecting the small target of the pests, and high feature similarity with the features between different categories. In enhancing the model, the partial convolution (PConv) module was employed to replace certain standard convolution modules in the original MobileViT module. Additionally, modifications were made to the feature fusion strategy within the MobileViT module, involving the concatenation fusion of input features, local expressive features, and global expressive features. The tenth layer MV2 module and the eleventh layer MobileViT module were removed, introducing an improved atrous spatial pyramid pooling (ASPP) module as a replacement, aiming to create multi-scale fusion features. Furthermore, the model adopted the SiLU activation function in lieu of the ReLU6 activation function for computations. Finally, the model underwent transfer learning based on the ImageNet dataset. The experimental results indicated that the recognition accuracy of six categories of fruit tree pest images reached 93.77%, with a parameter count of 8.40×105. In comparison with the previous version, the recognition accuracy was improved by 7.5 percentage points, while the parameter count was decreased by 33.86%. When compared with commonly used pest CNN recognition models, namely AlexNet, ResNet 50, MobileNetV2, and ShuffleNetV2, the proposed model achieved higher recognition accuracy by 8.25, 4.78, 7.27 and 7.41 percentage points, respectively, with parameter counts lowered by 6.03×107, 2.48×107, 2.66×106 and 5.30×105, respectively. Compared with Transformer recognition models such as ViT and Swin Transformer, the accuracy was higher by 19.03 and 9.8 percentage points, respectively, with parameter counts lowered by 8.56×107 and 2.75×107. The research was suitable for deployment in environments with limited resources, such as mobile terminals, which can contribute to the effective identification and detection of small target pests in fruit trees amidst complex backgrounds.