Previous Articles     Next Articles

Aquarium fish target detection APP development based on TensorFlow

  

  1. (1. Key Laboratory of East China Sea Fishery Resources Exploitation & Utilization, Ministry of Agriculture and Rural Affairs，, East China Sea Fisheries Research Institute, Chinese Academy of Fishery Sciences, Shanghai 200090, China;
     2. College of Information, Shanghai Ocean University, sShanghai, 201306, China;
    3. Shanghai Junding Fishery Technology Co., Ltd., Shanghai 200090, China;
    4. Shanghai fFishery Networking Technology Co., Ltd., Shanghai 200434, China)
  • Online:2020-09-16 Published:2020-04-20

基于TensorFlow的水族馆鱼类目标检测APP开发

  

  1. （1 中国水产科学研究院东海水产研究所，农业农村部东海渔业资源开发利用重点实验室，上海 200090；
    2  上海海洋大学信息学院，上海 201306；
    3 上海峻鼎渔业科技有限公司，上海 200090；
    4  上海渔联网科技有限公司，上海200434）
  • 通讯作者: 樊伟（1971－），男，博士，研究员，研究方向：渔业遥感与地理信息。E-mail:dhyqzh@sh163.net
  • 作者简介:张胜茂(1976 — )，男，博士，副研究员，研究方向：渔业数据挖掘、鱼类图像分析等领域。E-mail: ryshengmao@126.com
  • 基金资助:
    国家自然科学基金重点项目（61936014）；国家重点研发计划（2019YFD0901405）；农村农业部渔业渔政管理局项目(17190020)；杨浦区人力资源和社会保障局博士后项目（杨人社〔2019〕45号）

PDF (PC)

Abstract: In recent years, deep learning has made a breakthrough in image recognition research, which has led to the rapid development of target detection technology. It can enhance visitors’' experience and improve the effect of science popularization by using target detection technology to develop aquarium fish target detection appAPP. In this paper, for 80 species of fish photographed at the aquarium, we usedfirstly, labelimge LabelImg software was used to mark the target of 80 kinds of fishes photographed in aquarium. A, and the marked target was exported to generate tfrecord data. Secondly, ssd_mobilenet_v1MobileNetV1 model was selected model for data training, and we obtained the fish target detection model was obtained through 200 000 times of iterative training. Finally, Two 2 interfaces and 12 classes were defined with TensorFlow multi-targets detection API call model, and Android system mobile APP was developed. After testing 1 620 pictures of 80 species of fish were tested, and, the accuracy rate was 92.59%. The program used tensorflow multi-target detection API to call the model. The average target detection time target was 40ms on Huawei MHA-AL00. Fish target detection app APP can realize fast fish recognition in aquarium and real-time detection of multiple fish targets. It ,can improve visitors’' visiting experience and help contribute to quantitative evaluation of science popularization.

Key words: aquarium, target detection, TensorFlow, APP

摘要： 近年来深度学习在图像识别研究中取得突破进展，带动了目标检测技术的快速发展。利用目标检测技术开发水族馆鱼类目标检测APP，可以增强游客参观体验，提升科普效果。针对水族馆拍摄的80种鱼类，首先，使用LabelImg软件进行目标标记，再利用标记的目标导出成tfrecord数据；其次，选择ssd_mobilenet_v1模型进行数据训练，通过20万次的迭代训练获取到鱼类目标检测模型；最后，利用TensorFlow多目标检测API调用模型，定义2个接口和12个类，开发出Anroid系统手机APP。经过80种鱼类1 620张图片测试，正确率为92.59%，华为MHA-AL00手机目标检测平均时间40 ms。使用鱼类目标检测APP，能实现水族馆鱼类快速识别、多鱼类目标实时检测，可提升游客的参观体验，辅助科普量化评价。

关键词: 水族馆, 目标检测, TensorFlow, APP