利用深度卷積神經網路辨識延繩釣漁獲中常見之魚種

Abstract

漁獲統計資料是海洋資源管理中必不可少的資訊，而統計資料通常是由海上觀察員或船員手動紀錄的，但這樣人工紀錄的方式非常耗時且極具主觀性，因此存在自動化收集與回報漁獲資料的需求。然而漁船甲板上時常充滿各式的雜物，使得自動化的收集與回報漁獲資料充滿挑戰性。近年來由於卷積神經網路 (convolutional neural networks, CNNs) 越來越受歡迎且廣泛被應用於各種複雜的機器視覺任務。因此本研究利用深度卷積神經網路自動辨識11種延繩釣漁船常捕獲之魚種／類，這些魚種／類包含長鰭鮪魚 (Thunnus alalunga)、大目鮪魚 (T. obesus)、黃鰭鮪魚 (T. albacares)、南方黑鮪 (T. maccoyii)、黑皮旗魚 (Makaira nigricans)、雨傘旗魚 (Istiophorus platypterus)、劍旗魚 (Xiphias gladius)與鬼頭刀 (Coryphaena hippurus)。本研究之自動魚種辨識模型使用四種不同的深度卷積神經網路架構，包含：VGG-16、ResNet-50、DenseNet-201與MobileNetV2，搭配center loss function 進行訓練。研究中卷積神經網路模型之準確率最高可達95.83%，而在圖像顯示卡 (Graphics Processing Unit, GPU) 與中央處理器(Central Processing Unit, CPU)最快之運行速度分別可達1.75與107.82毫秒∕影像。

Fish catch statistics reported by vessels are essential information for the management of marine resource. The statistics were conventionally recorded by observers or fishermen. Manual recording is time consuming and can be subjective; thus, there is a demand for automatic statistics collection and reporting. The decks of fishing vessels are usually full of miscellaneous items, making automatic reporting of the statistics challenging. In recent years, convolutional neural networks (CNNs) have become increasingly popular and been applied to solving complex machine vision tasks. This study proposed to automatically identify 11 species or types of fish harvested by longliners using deep CNNs. The species included albacore (Thunnus alalunga), bigeye tuna (T. obesus), yellowfin tuna (T. albacares), southern bluefin tuna (T. maccoyii), blue marlin (Makaira nigricans), Indo-Pacific sailfish (Istiophorus platypterus), swordfish (Xiphias gladius), and dolphin fish (Coryphaena hippurus). Four deep CNNs modified from architectures VGG-16, ResNet-50, DenseNet-201, and MobileNetV2 were trained to identify the species and types of the fish in images collected on longliners. Center loss function was also applied during training for improving the performance of the CNNs. The CNNs reached an accuracy of as high as 95.83% and required a processing time of as short as 1.75 ms using a GPU and 107.82ms using a CPU.