白帶魚是臺灣近海重要的撈捕和消費的主要魚類之一。臺灣常見的白帶魚品種有三種，目前在撈捕和消費現場要進行白帶魚的品種辨識，都需藉由專家經驗人工辨識，因此難以達到普及應用目的。本研究提出一個白帶魚特徵擷取方法，藉由眼睛中心至頭部輪廓的距離向量，來鑑識白帶魚品種。我們首先透過U-Net神經網路進行白帶魚偵測以及頭部分割，接著提取輪廓以及偵測魚眼，最後計算每隻白帶魚的形狀特徵。為了驗證此一特徵的魚種鑑別性，我們以機率神經網路、決策樹、支援向量機以及k近鄰分類器等四種分類器來進行魚種辨識。;Trichiurus lepturus is one of the main fishes that are harvested and consumed off the coast of Taiwan. There are three common species of Trichiurus lepturus in Taiwan. Currently, fish’s classification at the fishing and consumption sites, manually identified the species of Trichiurus lepturus by experts or experiencers is required. Therefore, it is difficult to achieve the popularization purpose. This research mentions about a feature extraction method for Trichiurus lepturus, which uses the distance vector from the center of the eye to the outline of the head, to identify the species of Trichiurus lepturus. First, we use the U-Net neural network to detect the Trichiurus lepturus and execute the head segmentation, then extract the contours and detect fish eyes, and calculate the shape features of each Trichiurus lepturus. In order to verify the "discriminative" of shape feature, we use four types of classifiers such as probability neural network, decision tree, support vector machine and k-nearest neighbor classifier to classify the fishes.
