基於深度學習之變韌鐵電子顯微影像中MA島的輕量化分割模型

Abstract

本研究提出一個基於深度學習之變韌鐵電子顯微影像中 MA 島的輕量化分割模型，用於在變韌鐵顯微影像中分割麻田散鐵-沃斯田鐵（MA）島。MA 島在預測變韌鐵的抗衝擊性能中起著關鍵作用。傳統的 MA 島評估依賴專家主觀的意見，限制了評估結果的準確性和一致性。因此本研究透過訓練 MA 島分割模型時時融入不同專家的意見，來實現更客觀和穩定的分割結果。而在實際應用中，處理大量高分辨率電子顯微鏡圖像需要的大量計算資源。因此，本研究將如何輕量化模型作為一個重要的研究方向。本系統利用骨幹網路 (backbone network) 從變韌鐵顯微影像中提取相關特徵，接著使用頭部網路 (head network) 進行 MA 島分割。本研究探索兩種不同的骨幹網路，HRNet 和 Lite-HRNet，並進行輕量化的改良以減少模型的複雜性和提高效率。另外對於兩種骨幹網路本研究皆使用 OCRNet 作為頭部網路。對於 HRNet 的輕量化，通過替換 HRNet 的基本構建塊-ResNet 塊為 ConvNeXt 塊，並刪除某些逐點 (pointwise) 卷積層來輕量化 HRNet，最後使用 ECA 模組增強 HRNet 的性能。在 Lite-HRNet 中，通過用 ECA 模組替換 SW 和 CRW 模組，來降低 Lite-HRNet 的參數量以及計算複雜度。實驗結果顯示改良後的 HRNet 與 OCRNet 的組合相較於原始的 HRNet，參數數量和 FLOPs 分別減少 64% 和 27%，同時 MA 島 IoU 達到 78.14%。改良後的 Lite-HRNet 與 OCRNet 的組合相較於原始的 Lite-HRNet，參數數量和 FLOPs 分別減少 4.6% 和 11.66%，同時保持 MA 島 IoU 為 78.09%。由實驗結果發現，本研究所提出的改良方案，在輕量化的同時仍能保持變韌鐵電子顯微影像中 MA 島的分割準確性。

This study presents a deep-learning-based system for segmenting martensite–austenite (MA) islands inbainite electron micrographs. MA islands play a crucial role in predicting the impact resistance of bainite, but conventional assessment of them relies on subjective expert knowledge, leading to limitations in accuracy and consistency. To overcome this, the proposed system integrates expert opinions in order to achieve a more objective and stable segmentation result. However, in practical applications, a significant amount of computational resources is required to process the large number of high-resolution micrographs. Therefore, this study considers developing a lightweight model to be an important research direction.The proposed system utilizes a backbone network to extract relevant features from bainite micrographs, followed by a head network for MA island segmentation. This study explores two different backbone networks, High-Resolution Net (HRNet) and Lite-HRNet, and makes improvements to reduce model complexity and increase efficiency. In addition, OCRNet is selected as the head network for both backbone networks. To develop a lightweight of HRNet, its fundamental building blocks are replaced with ConvNeXt blocks and certain pointwise-convolution layers are removed to reduce the model’s complexity. Additionally, efficient channel attention (ECA) modules are used to enhance the performance of HRNet. In Lite-HRNet, the spatial weight and cross-resolution weight (CRW) modules are replaced with ECA modules to reduce the number of parameters and the computational complexity.The experimental results demonstrate that the combination of the improved HRNet and OCRNet achieves a reduction of 64% in number of parameters and 27% in floating-point operations (FLOP) compared to the original HRNet, while achieving an intersection over union (IoU) for MA islands of 78.14%. Similarly, the combination of the improved Lite-HRNet and OCRNet achieves a reduction of 4.6% in number of parameters and 11.66% in FLOPs compared to the original Lite-HRNet, while maintaining an MA island IoU of 78.09%. From the experimental results, it can be observed that the proposed improvement successfully achieves segmentation accuracy while reducing computational complexity.