首页分享基于CA

基于CA

来源：花匠小妙招时间：2024-12-25 06:36

摘要：

病害侵袭是制约核桃优质发展的重要因素之一,为实现田间智能化病害识别,设计了一款核桃病害识别模型。该模型采用Mobilenet-V2作为基础网络骨架,在倒残差结构中添加坐标注意力机制,解决特征提取时位置信息缺失的问题。此外,设计混合迁移的训练方式,将跨域迁移和域内迁移相结合,避免单独迁移学习的不良影响。结果表明:1)混合迁移对模型提升效果最佳,准确率最高提升18.57百分点。2)模型平均识别准确率为96.97%,模型参数量为3.95 M,内存占有量为10.50 MB,相较于Mobilenet-V3-Large、ShuffulNet-V2和EfficientNet-V2-S,识别准确率分别提升4.39百分点、6.63百分点和4.31百分点,且保持较少的参数量与内存占有量。3)与SE(squeeze-and-excitation)模块、CBAM(convolutional block attention module)模块相比,坐标注意力机制更能提升模型对感兴趣区域的关注度。因此,该模型可用于开发安卓应用程序并部署于移动端,为核桃病害智能识别提供新方法。

关键词: 核桃病害, 坐标注意力机制, 混合迁移, 安卓应用程序

Abstract:

Disease invasion is one of the important factors restricting the high-quality development of walnut. In order to realize intelligent disease identification in the field, this study designed a walnut disease identification model. The model used Mobilenet-V2 as the basic network skeleton, and added a coordinate attention mechanism to the inverted residual structure to make up for the lack of location information during feature extraction. In addition, this study designed a mixed transfer training method, which combined cross-domain and intra-domain to avoid the adverse effects of separate transfer learning. The results showed that: 1) The mixed transfer had the best effect on improving the model, and the highest accuracy rate was increased by 18.57 percentage points. 2) The average identification accuracy of the model was 96.97%, the model parameter size was 3.95 M, and the memory occupancy was 10.50 MB. Compared with Mobilenet-V3-Large, ShuffulNet-V2 and EfficientNet-V2-S, the identification accuracy was increased by 4.39, 6.63 and 4.31 percentage points, respectively, and the parameter size and memory occupation were keep small. 3) Compared with SE (squeeze-and-excitation) and CBAM (convolutional block attention module), the coordinate attention mechanism could improve the model’s attention to the region of interest. Therefore, the model could be used to develop an Android application and deploy it on the mobile terminal, and provide a new method for intelligent identification of walnut disease.

Key words: walnut disease, coordinate attention mechanism, mixed transfer, Android application

中图分类号:

S436.6