融合高级与低级视觉特征的农业图像显著性区域预测方法研究

Integration of high-and low-level visual features for salient region prediction of agricultural images

【目的】构建融合高级与低级视觉特征的农业图像(果实、农作物及畜禽目标)显著性区域预测算法,为农作物生长状态的监测、动物的体况评估提供支持。【方法】提出一种整合高级和低级视觉特征的农业图像显著性区域预测深度学习框架及其预训练方案。在MSRA10k数据集上按照6∶2∶2的比例进行训练、验证和测试,并采用F-Measure作为评价指标,在6种公共数据集(SOD、ASD、SED2、ECSSD、HKU-IS和THUR)及农业图像典型数据集上,将预测算法与4种显著性预测算法(MWS、IMS、FSN、P-Net)进行对比。【结果】所建立的预测算法在6种公共数据集上的平均F-Measure分数最高,为0.823,平均MAE分数最低,为0.099,显著性可视化结果边界完整,与人工标记的基准图像更接近。在农业图像典型数据集上的平均F-Measure为0.826,表明该算法可有效应对复杂农业场景的干扰,实现更为准确的目标轮廓信息提取。【结论】融合高级与低级视觉特征的图像显著性区域预测算法,可以实现对复杂农业场景下农作物及畜禽图像显著性区域的快速、准确预测。

【Objective】This study constructed a salient area prediction algorithm for agricultural images(fruits,crops and livestock targets)with integration of high-and low-level visual features to provide support for monitoring of crop growth and assessment of animal conditions.【Method】A deep learning framework was proposed for salient region prediction using the integration of high-and low-level visual features together with a corresponding pre-trained scheme.Verification and testing were carried out using the MSRA10k data set according to the ratio of 6∶2∶2.Using F-Measure as the evaluation index,the established prediction algorithms and 4 salient prediction algorithms(MWS,IMS,FSN and P-Net)were evaluated based on 6 public data sets(SOD,ASD,SED2,ECSSD,HKU-IS and THUR)and typical agricultural image data sets.【Result】The established prediction method had the highest average F-Measure of 0.823 and the lowest average MAE of 0.099 based on the 6 public data sets,and the salient visualization result had a complete boundary and was closer to the artificially marked reference images.The average F-Measure of established prediction algorithm on the typical agricultural image data set was 0.826,indicating that the algorithm could effectively deal with the interference of complex agricultural scenes and achieve more accurate target contour information extraction.【Conclusion】The image salient region prediction algorithm combing high-and low-level visual features could realize rapid and accurate prediction of salient region from crops and livestock images in complex agricultural scenes.