基于增强型Tiny-YOLOV3模型的野鸡识别方法

智慧农业病虫害检测技术发展迅猛，而对农作物具有危害的鸟类检测技术尚处于起步阶段，近年来由于生态改善，野鸡繁殖数量激增，其喜食小麦、玉米、红薯等农作物的种子与幼苗，对农业造成一定危害。该研究提出了一种适宜于嵌入式系统部署的人工智能野鸡识别方法。由于在野外环境下移动平台上部署，需采用轻量级网络，同时保证检测精度与实时性，因此，根据Tiny-YOLOV3轻量级目标检测网络基本结构，提出了一种针对野外复杂环境中出现野鸡的实时检测网络-增强型轻量级目标检测网络（Enhanced Tiny-YOLO，ET-YOLO），该网络特征提取部分加深Tiny-YOLOV3特征提取网络深度，增加检测尺度以提高原网络目标检测精度，网络检测层使用基于CenterNet结构的检测方式以进一步提高检测精度与检测速度。使用野外实地采集各种环境下出现的野鸡图像作为数据集，包括不同距离、角度、环境出现的野鸡共计6 000幅高清图像制作数据集。试验结果表明，ET-YOLO在视频中复杂环境下出现的野鸡平均检测精度达86.5%，平均检测速度62帧/s，相对改进前Tiny-YOLOV3平均检测精度提高15个百分点，平均检测速度相对改进前Tiny-YOLOV3提高2帧/s，相对YOLOV3、Faster-RCNN与SSD_MobileNetV2主流代表性目标检测算法，平均检测精度分别提高1.5、1.1与18个百分点，平均检测速度分别提高38、47与1帧/s。可高效实时地对复杂环境下出现的野鸡进行识别，并且检测模型大小为56 MB，适宜于在农业机器人，智能农机所搭载的嵌入式系统上部署。

Recognition method of pheasant using enhanced Tiny-YOLOV3 model

The increase of pheasants has posed a threaten to crops as the advancement of ecology. However, most conventional methods of bird repellent have inherent deficiencies in terms of efficiency and danger. An efficiency monitoring method for pheasant is necessary to combine with artificial intelligence, in order to provide early warning and expulsion of pheasants. Normally, pheasant activities are mostly in the early morning and dusk under complex environment with protective color or habit of hiding. This behavior has made monitoring methods much more challenge. In this paper, a novel recognition method for pheasant has been proposed on the deployment of embedded computing platform, combined with the enhanced Tiny-YOLOV3 target detection network, particularly on considering the behavior of pheasant and specific living conditions. A lightweight network is required to ensure the accuracy and real-time monitoring due to the deployment on a mobile platform in the field environment. A real-time monitoring network ET-YOLO has also been established for the emergence of pheasants in a complex field environment, according to the basic structure of the Tiny-YOLOV3 lightweight target detection network. The feature extraction can deepen the net depth of Tiny-YOLOV3, and thereby increase the detection scale to improve the detection accuracy of original net target. CenterNet structure was used in the net detection layer to further enhance the detection accuracy and speed. The dataset of pheasant monitoring was produced after augmentation using the field collection of images in various environments, including 6000 high resolution images of pheasant in different distances, angles and environments. The indicators of experimental evaluation were mainly tested in terms of accuracy, real-time performance, and model size. Specifically, the average detection accuracy, average detection speed, and detection model size of the pheasant were used for evaluation. The experimental results showed that the average detection accuracy of ET-YOLO in the complex field environment was 86.5%, and the average detection speed was 62 frames/s, 15% higher than that of initial Tiny-YOLOV3. The average detection accuracy was higher than that of YOLOV3, Faster-RCNN and SSD_MobileNetV2 by 1.5%, 1.1% and 18%, respectively. The average detection speed was 38 frames /s, 47 frames /s and 1 frame/s higher than that of YOLOV3, Faster-RCNN and SSD_MobileNetV2, respectively, when the detection model size of 56 MB. The proposed method can be suitable for the deployment on embedded computing platforms equipped with agricultural robots and intelligent machines in terms of recognition accuracy, real-time performance, and model size, particularly recognizing pheasants in complex environments.