郭庆文, 王春桃, 肖德琴, 黄琼. 利用显著图构建注意力深度网络检测诱虫板蔬菜害虫[J]. 农业工程学报, 2021, 37(19): 211-219. DOI: 10.11975/j.issn.1002-6819.2021.19.024
    引用本文: 郭庆文, 王春桃, 肖德琴, 黄琼. 利用显著图构建注意力深度网络检测诱虫板蔬菜害虫[J]. 农业工程学报, 2021, 37(19): 211-219. DOI: 10.11975/j.issn.1002-6819.2021.19.024
    Guo Qingwen, Wang Chuntao, Xiao Deqin, Huang Qiong. Building saliency-map-based attention-driven deep network to detect vegetable pests of sticky trap images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(19): 211-219. DOI: 10.11975/j.issn.1002-6819.2021.19.024
    Citation: Guo Qingwen, Wang Chuntao, Xiao Deqin, Huang Qiong. Building saliency-map-based attention-driven deep network to detect vegetable pests of sticky trap images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(19): 211-219. DOI: 10.11975/j.issn.1002-6819.2021.19.024

    利用显著图构建注意力深度网络检测诱虫板蔬菜害虫

    Building saliency-map-based attention-driven deep network to detect vegetable pests of sticky trap images

    • 摘要: 为提高诱虫板图像蔬菜害虫检测精度,针对背景区域容易导致误检的问题基于显著图分析技术构建了一种注意力深度网络害虫智能视觉检测方法。首先通过显著图筛选出粗候选区域;然后在粗候选区域内用全卷积神经网络精选出细候选区域;接着用神经网络分类器识别细候选区域害虫种类,得到含有冗余的若干检测框;最后用改进的非极大值抑制消除冗余检测框,实现诱虫板图像中目标害虫的检测。针对小菜蛾和瓜实蝇展开试验,获得86.40%的平均精度均值和0.111只的平均绝对计数误差均值,所提方法平均精度均值比Faster R-CNN和YOLOv4分别高2.74和1.56个百分点,平均绝对计数误差均值比Faster R-CNN和YOLOv4分别低0.006和0.003只;同时,消融试验中移除显著图注意力模块后平均精度均值下降了4个百分点、平均绝对计数误差均值增加了0.207只。试验结果表明,所提方法有效提高了诱虫板图像蔬菜害虫检测精度,其中,引入显著图注意力模块对提升检测精度有重要作用。

       

      Abstract: Digital imaging has widely been used to detect pest diseases for crops in modern agriculture, particularly on deep learning and intelligent computer vision. However, accurate and rapid detection of insect pests in images still remains a great challenge in the crop field. In this study, a task-specified detector was developed to accurately detect vegetable pests of sticky trap images using an attention-driven deep network from saliency maps. Prevailing pest detectors were mainly adopted anchors to detect pests in sticky trap images. Nevertheless, the anchor-based detection accuracy depended mainly on the balance between positives and negatives, as well as the model training, due mainly to the relatively small sizes and distribution of crop insect pests in the sticky trap images. Therefore, a saliency map was established to filter simple background regions. An attention-driven neural network was also selected to better focus on key regions and then accurately detect crop insect pests of sticky trap images. Firstly, saliency maps and threshold-based techniques were employed to construct masks for rough region proposals, according to connected graphs of acquired masks. Secondly, two fully convolutional neural networks were used in a sliding window fashion to produce refined region proposals from rough region proposals, in order to deal with occlusion issues. Thirdly, each refined region proposal was then classified as one target pest category with a convolutional neural network classifier, thereby detecting the bounding boxes of target vegetable pests. Finally, an enhanced non-maximum suppression was utilized to eliminate the bounding boxes of redundant detection, where a target pest was captured by only one detection bounding box. As such, the target pest number was easily obtained to count the bounding boxes of rest detection during automatic management of vegetable insect pests. Furthermore, a piece of specific monitoring equipment was designed to evaluate the vegetable pest detector, where sticky trap images of two vegetable pests were collected, including Plutellaxylostella (Linnaeus) and Bactroceracucuribitae (Coquillett). Several experiments were also conducted on the labeled data set of collected images. The results demonstrate that the vegetable pest detector achieved a mean average precision of 86.40% and an average mean absolute error of 0.111, indicating better performance than the commonly-used pest detectors, such as SSD, R-FCN, CenterNet, Faster R-CNN, and YOLOv4. In addition, two ablation experiments were carried out to verify the attention mechanism of saliency maps and the enhanced non-maximum suppression. It was found that the attention mechanism remarkably contributed to the detection accuracy and the performance of enhanced non-maximum suppression. In the future, both top- and low-level feature maps were required in a convolutional neural network, further enchancing the robustness of the attention mechanism in the vegetable pest detector.

       

    /

    返回文章
    返回