Energy-Efficient Image Recognition System for Marine Life
Energy-Efficient Image Recognition System for Marine Life
复制标题
DOI：
10.1109/tcad.2020.3012745
复制
发表时间：
2020-11
影响因子：
2.9
通讯作者：
H. S. Demir;J. Christen;S. Ozev
中科院分区：
文献类型：
--
作者：
H. S. Demir;J. Christen;S. Ozev
本文重点设计了一种适用于海洋监测的节能图像识别系统。水下成像系统的主要挑战之一是由于现场资源有限而导致的严格的功耗约束。考虑到在不同的水浊度水平和背景照明条件下连续运行的需要，需要一种节能的方法来有效地利用资源。在这项工作中，我们提出了一个识别框架，将自适应地调整系统参数，如相机帧速率和LED照明水平，根据环境条件，以优化能源消耗，同时确保高识别精度。所提出的决策系统的第一部分包含基于卷积神经网络（CNN）的动物识别块，其用于获得单个帧的置信水平。第二部分是自适应决策块，它动态地改变系统参数，并根据环境条件组合多帧识别块的结果。在我们的实验中，我们已经使用了近8000水下图像的训练和测试的单帧识别块，并使用了近200个不同的视频序列的训练和测试的自适应决策块。基于由Raspberry Pi 3 Model B、Pi NoIR Camera v2.1和850 nm LED组成的硬件框架的测量，所提出的系统通过基于水浊度和背景照明水平动态改变帧速率和发射光强度，实现了高达92.7%的节能，并具有相当的识别性能。
This article focuses on designing an energy-efficient image recognition system for marine monitoring. One of the main challenges of an underwater imaging system is the strict power consumption constraints due to the limited on-site resources. Considering the need for continuous operation in different water turbidity levels and background illumination conditions, an energy-efficient approach is needed for the effective utilization of the resources. In this work, we propose a recognition framework that will adaptively adjust the system parameters, such as camera frame rate and LED illumination level, based on the environmental conditions to optimize the energy consumption while ensuring a high recognition accuracy. The first part of the proposed decision system contains the convolutional neural network (CNN)-based animal recognition block which is used for obtaining the confidence level for a single frame. The second part is the adaptive decision block that dynamically changes the system parameters and combines the results of the recognition block for multiple frames based on the environmental conditions. In our experiments, we have used nearly 8000 underwater images for training and testing the single frame recognition block and used nearly 200 different video sequences for training and testing the adaptive decision block. Based on measurements of a hardware framework composed of a Raspberry Pi 3 Model B, a Pi NoIR Camera v2.1, and 850 nm LEDs, the proposed system achieves up to 92.7% energy savings with a comparable recognition performance by dynamically changing the frame rate and emitted light intensity based on water turbidity and background illumination level.