agriculture Article Moth Detection from Pheromone Trap Images Using Deep Learning Object Detectors Suk-Ju Hong 1 , Sang-Yeon Kim 1 , Eungchan Kim 1, Chang-Hyup Lee 1, Jung-Sup Lee 2, Dong-Soo Lee 3, Jiwoong Bang 2 and Ghiseok Kim 1,4,* 1 Department of Biosystems and Biomaterials Science and Engineering, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Korea; [email protected] (S.-J.H.); [email protected] (S.-Y.K.); [email protected] (E.K.); [email protected] (C.-H.L.) 2 Protected Horticulture Research Institute, National Institute of Horticultural and Herbal Science, Rural Development Administration, 1425, Jinham-ro, Haman-myeon, Haman-gun, Gyeongsangnam-do 52054, Korea; [email protected] (J.-S.L.); [email protected] (J.B.) 3 Department of Leaders in Industry-University Cooperation, Chung-Ang University, 4726, Seodong-daero, Daedeok-myeon, Anseong-si, Gyeonggi-do 17546, Korea; [email protected] 4 Research Institute of Agriculture and Life Sciences, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Korea * Correspondence: [email protected]; Tel.: +82-2-880-4603 Received: 7 April 2020; Accepted: 12 May 2020; Published: 14 May 2020 Abstract: Diverse pheromones and pheromone-based traps, as well as images acquired from insects captured by pheromone-based traps, have been studied and developed to monitor the presence and abundance of pests and to protect plants. The purpose of this study is to construct models that detect three species of pest moths in pheromone trap images using deep learning object detection methods and compare their speed and accuracy. Moth images in pheromone traps were collected for training and evaluation of deep learning detectors. Collected images were then subjected to a labeling process that defines the ground truths of target objects for their box locations and classes. Because there were a few negative objects in the dataset, non-target insects were labeled as unknown class and images of non-target insects were added to the dataset. Moreover, data augmentation methods were applied to the training process, and parameters of detectors that were pre-trained with the COCO dataset were used as initial parameter values. Seven detectors—Faster R-CNN ResNet 101, Faster R-CNN ResNet 50, Faster R-CNN Inception v.2, R-FCN ResNet 101, Retinanet ResNet 50, Retinanet Mobile v.2, and SSD Inception v.2 were trained and evaluated. Faster R-CNN ResNet 101 detector exhibited the highest accuracy (mAP as 90.25), and seven different detector types showed different accuracy and speed. Furthermore, when unexpected insects were included in the collected images, a four-class detector with an unknown class (non-target insect) showed lower detection error than a three-class detector. Keywords: pheromone trap; pest; moth; deep learning; horticulture; insect detection 1. Introduction Continuous monitoring of the pest population is one of the essential elements in the pheromone-based pest control system [1]. Traditional monitoring of pheromone-based traps was conducted by manually identifying the species and population of pests on the trap. This approach is labor intensive and time consuming, and requires skilled personnel capable of distinguishing different species of pests. These disadvantages hinder pest monitoring in real time operation which can guarantee a specified timing constraint. In particular, early diagnosis of pests, one of the major challenges in the horticulture industry, requires automatic monitoring rather than manual monitoring [2]. Agriculture 2020, 10, 170; doi:10.3390/agriculture10050170 www.mdpi.com/journal/agriculture Agriculture 2020, 10, 170 2 of 12 Recently, research applying IoT (Internet of Things) technologies for farm automation and real-time farm monitoring have been widely conducted [3–5], and the applications of IoT technologies in agriculture enabled real-time remote monitoring of farms using vision systems. Among various vision systems, real-time monitoring equipment using an RGB (Red, Green, and Blue) color camera is fairly configurable because of its relatively inexpensive cost, and it can obtain useful information including shapes, colors and textures. In particular, owing to the rapid development of image processing technology, the effectiveness of RGB image-based monitoring has been increasing in recent years. Therefore, image-based monitoring techniques are being applied in many areas of farm monitoring, such as crop disease monitoring [6,7] and crop growing monitoring [8,9] systems. Traditional image classification or detection studies generally used image processing methods like active contour [10], scale-invariant feature transform (SIFT) [11], Histogram of Oriented Gradients (HOG) [12], Haar-like features [13] and machine learning methods like support vector machine (SVM) [14], artificial neural network (ANN) [15]. Image-based pest monitoring studies mainly used traditional image processing and machine learning techniques. Xia et al. [16] separated pests and background using watershed segmentation and detected pests by classifying the species of pests using a Mahalanobis distance. Li et al. [17] detected small-sized pests on leaf surfaces by applying multifractal analysis. Wen et al. [18] used a scale-invariant feature transform (SIFT) descriptor and six classifiers, including the minimum least square linear classifier (MLSLC) and K nearest neighbor classifier (KNNC) to classify orchard insects. Wang et al. [19] classified insect images using a support vector machine (SVM) and artificial neural network (ANN) after feature extraction. Bakkay et al. [20] detected moths using adaptive k-means clustering approach by using contour’s convex hull and region merging algorithm. Solis-Sanchez et al. [21] applied SIFT descriptor and LOSS algorithm to detect insects on trap. Bechar et al. [22] used pattern recognition methods using local maximality to detect whiteflies. The detection or classification, using traditional methods, was carried out by a person specifying features manually, and applying the machine learning algorithm after extracting features. However, in order to detect or classify images of various environments, it was necessary to combine not only simple features but also various features that humans cannot understand. Traditional methods of feature extraction and machine learning have made it impossible to create and apply these complex features. Recently, deep-learning methods, based on convolutional neural networks (CNNs), have mainly been used in the field of image classification and detection. In the case of CNN, the biggest difference from traditional methods is that it finds the characteristics and suitable features of target object for a given problem by learning data by itself. If the limitation of the existing methods is to use the features set by humans, CNN can learn and apply complexly from simple features to complex features that are difficult for humans to understand through deep layers. R-CNN, the first to apply CNN for object detection, outperformed existing object detection models based on feature extraction and machine learning [23]. Faster R-CNN, developed from R-CNN, increases the detection speed by applying a region proposal network (RPN), enabling real-time deep-learning-based detection. Faster R-CNN is 250 times faster than R-CNN and 10 times faster than Fast R-CNN [24], with higher accuracy [25]. After two-stage R-CNN type object detectors, one-stage object detectors, such as you only look once (YOLO) [26], single shot multibox detectors (SSD) [27], and Retinanet [28] were developed. These one-stage detectors showed significantly faster speed than two-stage detectors. Among them, Retinanet showed a high accuracy similar to the two-stage detector and the accuracy of small object detection, which was one of the weak points of the one-stage detector, was also superior to YOLO and SSD. Deep learning-based object detectors are widely used in vision fields such as pedestrian detection [29], crop disease detection [30], and object detection in aerial images [31,32]. In addition, it is commonly known that detectors based on these deep learning-based detection methods overwhelms the performance of traditional detection methods in most fields of detection application without the traditional labor-intensive feature engineering process. Therefore, in the field of pest detection Agriculture 2020, 10, 170 3 of 12 Agriculturefrom pheromone 2020, 10, x FOR trap PEER images, REVIEW studies applying deep-learning methods are increasingly required3 of 13 to accurately detect and classify pests having various characteristics for various environments. In recent environments.years, deep learning-based In recent years, methods deep learning-based have been applied methods to pest have detection been applied in pheromone to pest detection trap images. in pheromoneDing and Taylor trap [images.33] detected Ding moths and Taylor using a[33] sliding detected window moths method using and a sliding CNN. Sunwindow et al. method [34] detected and CNN.a red turpentineSun et al. beetle[34] detected (RTB) using a red a turpentine deep learning-based beetle (RTB) one-stage using objecta deep detector. learning-based Although one-stage research objecthas been detector. conducted Although on pestresearch detection has been methods conduc inted pheromone on pest detection traps for methods real-time in pheromone pest monitoring, traps forit is real-time necessary pest to studymonitoring, the optimal it is necessary method considering to study the various optimal deep method learning-based
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages12 Page
-
File Size-