基于轻量化深度学习Mobilenet-SSD网络模型的海珍品检测方法

为精确掌握水下海珍品养殖分布情况,摆脱传统上依赖人工潜水了解海珍品情况的方式,提出了一种基于轻量化深度学习的Mobilenet-SSD网络模型并用于海珍品检测,该方法对在渔船下方的水下摄像头所采集的海珍品图像实时进行目标快速检测。结果表明:采用本研究中建立的Mobilenet-SSD模型,在海胆、海参、扇贝等3种海珍品上建立数据集进行训练,可实现水下海珍品的精确识别,海胆、海参、扇贝的识别准确率分别为81.43%、86.02%、89.44%,总体平均准确率为85.79%;将Mobilenet-SSD网络模型分别与Tiny-YOLO和VGG-SSD网络模型进行比较,在相同设备上,Mobilenet-SSD网络模型相较Tiny-YOLO网络模型能更好地利用目标特征,同时较VGG-SSD网络模型节约80%的用时,实现了准确性与实时性的兼顾。研究表明,本研究中构建的Mobilenet-SSD网络模型,可用于水产养殖环境中水下海珍品的准确识别。

Detection method of high value marine food organisms based on lightweight deep learning Mobilenet-SSD network

In order to accurately understanding of the culture distribution of high value sea food organisms in water and get rid of the traditional way of relying on diving to probe into the situation of marine sea organisms,a lightweight deep learning Mobilenet-SSD network based detection method of high value sea food organisms was proposed via collecting real-time images of marine organisms through underwater cameras under fishing boats and quickly detect target marine organisms.The self-built data set was trained on Mobilenet-SSD network to realize the accurate recognition of 3 types of high value marine food organisms,with recognition rate of 81.43%in sea urchin,86.02%in sea cucumber and 89.44%in scallop,and average accuracy of 85.79%in the test.The comparison of Mobilenet-SSD with Tiny-YOLO network and VGG-SSD network on the same device,repectively,indicated that Mobilenet-SSD in the case of no loss of accuracy had more both accuracy and real-time than Tiny-YOLO network did,with saving 80%of time compared with VGG-SSD network.