欢迎您访问《智慧农业(中英文)》官方网站! English

Smart Agriculture ›› 2022, Vol. 4 ›› Issue (1): 140-149.doi: 10.12133/j.smartag.SA202203003

• 信息处理与决策 • 上一篇    

农业复杂环境下尺度自适应小目标识别算法——以蜜蜂为研究对象

郭秀明(), 诸叶平, 李世娟, 张杰, 吕纯阳, 刘升平()   

  1. 中国农业科学院农业信息研究所/农业农村部农业信息服务技术重点实验室,北京 100081
  • 收稿日期:2021-12-31 出版日期:2022-03-30
  • 基金资助:
    中央级公益性科研院所基本科研业务费专项(2021JKY038);河北省重点研发计划项目子课题(19227407D);国家蜂产业技术体系专项(CARS-44)
  • 作者简介:郭秀明(1981-),女,博士,副研究员,研究方向为农业智能感知。E-mail:guoxiuming@caas.cn
  • 通信作者:

Scale Adaptive Small Objects Detection Method in Complex Agricultural Environment: Taking Bees as Research Object

GUO Xiuming(), ZHU Yeping, LI Shijuan, ZHANG Jie, LYU Chunyang, LIU Shengping()   

  1. Agricultural Information Institute, Chinese Academy of Agricultural Sciences/Key Laboratory of Agri-information Service Technology, Ministry of Agriculture and Rural Affairs, Beijing 100081, China
  • Received:2021-12-31 Online:2022-03-30

摘要:

农业生产环境中的目标识别对象常具有分布密集、体积小、密度大的特点,加之农田环境光照多变、背景复杂,导致已有目标检测模型无法取得令人满意的效果。本研究以提高小目标的识别性能为目标,以蜜蜂识别为例,提出了一种农业复杂环境下尺度自适应小目标识别算法。算法克服了复杂多变的背景环境的影响及目标体积较小导致的特征提取困难,实现目标尺度无关的小目标识别。首先将原图拆分为一些较小尺寸的子图以提高目标尺度,将已标注的目标分配到拆分后的子图中,形成新的数据集,然后采用迁移学习的方法重新训练并生成新的目标识别模型。在模型的使用中,为使子图识别结果能正常还原,拆分的子图之间需具有一定的重叠率。收集所有子图的目标识别结果,采用非极大抑制(Non-Maximum Suppression,NMS)去除由于模型本身产生的冗余框,提出一种交小比非极大抑制(Intersection over Small NMS,IOS-NMS)进一步去除子图重叠区域中的冗余框。在子图像素尺寸分别为300×300、500×500和700×700,子图重叠率分别为0.2和0.05的情况下进行验证试验,结果表明:采用SSD(Single Shot MultiBox Detector)作为框架中的目标检测模型,新提出的尺度自适应算法的召回率和精度普遍高于SSD模型,最高分别提高了3.8%和2.6%,较原尺度的YOLOv3模型也有一定的提升。为进一步验证算法在复杂背景中小目标识别的优越性,从网上爬取了不同尺度、不同场景的农田复杂环境下的蜜蜂图像,并采用本算法和SSD模型进行了对比测试,结果表明:本算法能提高目标识别性能,具有较强的尺度适应性和泛化性。由于本算法对于单张图像需要多次向前推理,时效性不高,不适用于边缘计算。

关键词: 目标检测, 机器视觉, 小目标, 农业环境, 蜜蜂, SSD, YOLOv3

Abstract:

Objects in farmlands often have characteristic of small volume and high density with variable light and complex background, and the available object detection models could not get satisfactory recognition results. Taking bees as research objects, a method that could overcome the influence from the complex backgrounds, the difficulty in small object feature extraction was proposed, and a detection algorithm was created for small objects irrelevant to image size. Firstly, the original image was split into some smaller sub-images to increase the object scale, and the marked objects were assigned to the sub-images to produce a new dataset. Then, the model was trained again using transfer learning to get a new object detection model. A certain overlap rate was set between two adjacent sub-images in order to restore the objects. The objects from each sub-image was collected and then non-maximum suppression (NMS) was performed to delete the redundant detection boxes caused by the network, an improved NMS named intersection over small NMS (IOS-NMS) was then proposed to delete the redundant boxes caused by the overlap between adjacent sub-images. Validation tests were performed when sub-image size was set was 300×300, 500×500 and 700×700, the overlap rate was set as 0.2 and 0.05 respectively, and the results showed that when using single shot multibox detector (SSD) as the object detection model, the recall rate and precision was generally higher than that of SSD with the maximum difference 3.8% and 2.6%, respectively. In order to further verify the algorithm in small target recognition with complex background, three bee images with different scales and different scenarios were obtained from internet and test experiments were conducted using the new proposed algorithm and SSD. The results showed that the proposed algorithm could improve the performance of target detection and had strong scale adaptability and generalization. Besides, the new algorithm required multiple forward reasoning for a single image, so it was not time-efficient and was not suitable for edge calculation.

Key words: object detection, machine vision, small object, farmland, bee, single shot multibox detector, YOLOv3

中图分类号: