动态尺度特征挖掘与双向注意力交互的黄瓜病害检测

Cucumber Disease Detection with Dynamic Scale Feature Mining and Bidirectional Attention Interaction

摘要:

针对黄瓜病害检测中各类病斑难以区分和识别效率低等问题,提出动态尺度特征挖掘与双向注意力交互的黄瓜病害检测算法。在特征提取阶段引入感受野与通道注意力卷积RFCAConv,强化病斑区域的特征表征,提升模型对形态变异较大病害目标的识别能力;采用自适应特征增强模块AFEM,有效抑制背景噪声,增强目标与背景的区分度;设计C3k2-DSC模块,有效减少因目标形状变化导致的特征信息丢失。在构建的黄瓜病害数据集上进行消融试验,结果表明平均精度均值达到85.9%,相较于基线模型提升3.8个百分点,召回率提升5.1个百分点,各项改进均对模型识别性能有提升效果,其中,AFEM模块对模型提升效果最为明显;模型在包含3 070幅图像的公开数据集上试验结果表明,精确率和召回率分别提升2.3个百分点和4.4个百分点。将模型部署在Android手机端,自然环境下测试结果表明,平均识别准确率为84.8%,平均检测耗时为0.352 s。所建立的黄瓜病害检测系统能实现对黄瓜病害的精准快速识别,为黄瓜病害的及时防治提供有效方法。

Abstract:

Aiming to address the challenges of difficult-to-distinguish and inefficiently recognized disease spots in cucumber disease detection,a cucumber disease detection algorithm combining dynamic scale feature mining and bidirectional attention interaction was proposed. In the feature extraction stage,receptive field and channel attention convolution (RFCAConv) was introduced to enhance the feature representation of disease spot areas and improve the model's ability to recognize disease targets with significant morphological variations. An adaptive feature enhancement module (AFEM) was employed to effectively suppress background noise and enhance the distinction between targets and backgrounds. The C3k2-DSC module was designed to effectively reduce the loss of feature information caused by target shape variations. Ablation experiments conducted on a constructed cucumber disease dataset showed that the mean average precision reached 85.9%,an improvement of 3.8 percentage points compared with that of the baseline model,and the recall rate was increased by 5.1 percentage points. All improvements enhanced the model's recognition performance,with the AFEM module showing the most significant improvement. The model was tested on a public dataset containing 3 070 images,and the experimental results showed that the precision and recall rates were increased by 2.3 percentage points and 4.4 percentage points,respectively. When deployed on an Android mobile phone,the test results showed an average recognition accuracy of 84.8% and an average detection time of 0.352 s. The established cucumber disease detection system achieved accurate and rapid identification of cucumber diseases,providing an effective method for timely prevention and control of cucumber diseases.