本研究以安裝在延繩釣漁船上的電子觀察員所取得的定格相片為對象，使用Sobel濾波器擷取影像的特 徵換算成直方圖後，再分別以不同的判斷機制測試，嘗試找出有能力辨識該相片中是否有鮪魚的方法。擷 取的影像特徵包含:梯度強度、梯度角度與曲率等三種型式，而判斷機制選用相關係數為比較基準，採用 K-means分群法進行的測試，結果顯示辨識成效不如預期。提出「魚體全域特徵」及「魚體區域特徵」的 方式，先建構單一模板方法測試，證實有效性後再增加模板種類，並結合Bagging演算法，可以提升辨識的 正確率。測試結果得到96.0%有鮪魚的偵測率。此結果顯示搭配現場狀況，建構足夠模板數的話可以提高 辨識的正確率。

This study attempts to explore the methods to identify fishes in still images taken from the electronic observer system installed on longline fishing vessels. Sobel filter was used to capture image features and transform it into histogram. Different recognition mechanisms were tested to find out the best method. Captured image features include: gradient magnitude, angle, and curvature. Correlation coefficient was chosen as baseline of recognition mechanisms. K-means clustering method was used; however, the results showed the effectiveness of recognition is not as expected. Thus, “fish global feature” and “fish local feature” were proposed. First, a single tuna template was created and tested. After the validity was established, by combining more templates and utilizing Bagging algorithm, the recognition accuracy can be improved. In the experiment, 96.0% fish recognition rate was obtained. The results show that adequate number of templates were created with corresponding with site conditions, the recognition accuracy can be increased.