在監督式學習的分類問題中，一旦有新類別加入模型就可能會發生災難性遺忘的問題且無法新增未知樣本至辨識系統當中，通常分類模型必須重新訓練才能解決此問題。本篇論文提出一個基於對比式學習與三元組網路架構的特徵擷取模型，稱為多層特徵卷積神經網路(Multi-Level Feature Convolution Neural Network, MLF-CNN)，可以同時擷取低階特徵與高階特徵，我們以此設計了一個可擴增類別的魚類辨識系統，讓魚類辨識模型能夠擴增未知魚種類別。我們同時提出四個步驟的解釋程序，來分析辨識系統分類錯誤的原因。實驗結果顯示ResNet34與MLF-CNN在已知魚種分類任務的Top-1準確度皆為98%；在未知魚種類別擴增任務中，MLF-CNN準確度優於ResNet34約2%；整體系統效能的評估測試中，MLF-CNN略優於ResNet34約0.05%。證明採用MLF-CNN的魚類辨識系統能有更好的系統效能表現。;In supervised learning classification problems, once a new class is added to the model, the problem of catastrophic forgetting and the inability to add unknown samples to the recognition system may occur, and the classification model usually has to be retrained to solve this problem. In this paper, we propose a feature acquisition model based on contrast learning and triplet network architecture, called Multi-Level Feature Convolution Neural Network(MLF-CNN), which can acquire both low-level features and high-level features. We also propose a four-step interpretation procedure. We also propose a four-step explanation procedure to analyze the causes of misclassification in the recognition system. The experimental results show that both ResNet34 and MLF-CNN have 98% accuracy in Top-1 for the known species classification task; MLF-CNN has better accuracy than ResNet34 by about 2% in the unknown species augmentation task; MLF-CNN is slightly better than ResNet34 by about 0.05% in the overall system performance evaluation. This demonstrates that the fish identification system using MLF-CNN can have better system performance.
