Authors

Abstract

The real-time images captured by agricultural machinery on-board monitoring equipment have complex backgrounds and different shooting angles. Especially for straw monitoring tasks, the utilization rate of images is relatively low. This paper presents a novel image classification and effective region segmentation method for straw returning in agriculture, leveraging semantic segmentation to enhance the efficiency of agricultural data analysis. The study addresses the challenges of manual straw cover analysis by proposing an automated approach to select images that meet monitoring standards. The methodology employs an encoder-decoder structure model, enriched with residual units, multi-scale convolution, and attention mechanisms. This model classifies images by calculating the pixel proportions of various scene categories and segments farmland areas to be inspected by incorporating distance information. The model's design is tailored to handle the complex and variable natural environments typical of vehicular monitoring scenarios, where semantic object boundaries can be fuzzy. The experimental results demonstrate that the proposed method achieves an overall sample classification accuracy of 93% for straw returning image classification and an 85.37% accuracy in dividing areas to be inspected. The method outperforms several mainstream semantic segmentation models, providing a more accurate and efficient means of processing agricultural monitoring images. The integration of distance information proves particularly beneficial in distinguishing the farmland areas under inspection, leading to clearer segmentation and more reliable data for agricultural decision-making. In conclusion, the study contributes to the field of agricultural intelligence by offering a robust method for image analysis that can be applied to optimize the use of straw return monitoring data.

Abstract in Chinese

农机车载监控设备拍摄的实时图像背景复杂，特别是对于秸秆定量检测任务，图像的利用率相对较低。本文提出了一种基于语义分割的图像高效分类和背景区域分割方法，通过计算各种场景类别的像素比例对图像进行分类，筛选出符合秸秆检测要求的图像，再并结合距离信息进一步分割秸秆检测区域，提高车载图像数据分析利用效率与秸秆检测准确度，以此来解决农田大数据杂乱，利用率低的问题。该方法采用编码器-解码器模型结构，融合了轻量型残差单元、多尺度卷积和注意力机制，在保证分割边界清晰的情况下降低模型参数。从不同模型的对比结果和可视化处理结果可以表明：该方法对秸秆图像分类的总体样本分类准确率为93%，对待检区域的划分准确率为85.37%，该模型对背景类别中农田的分割效果更好，空间位置更加准确。因此，本文的研究方法能提高秸秆信息化监测手段的应用效率，为农田监测图像的处理提供了一种更准确、更高效的方法。