基於循環收縮聚合技術之語意分割

Abstract

目前最先進的語意分割技術模型，其構造往往結合了卷積神經網路(convolutional neural network)以及條件隨機域(conditional random field)。作為前端的初始語意分類器，卷積神經網路產生的預測結果往往含有邊界錯誤。另一方面，作為優化的系統，條件隨機域藉著強調區域標籤的一致性，修正前端所產生之語意分割結果中的錯誤，往往能有效改善語意分割的結果。然而，條件隨機域的一大缺陷為其改善結果時常伴隨邊界模糊化現象，基於此原因，我們傾向使用別的方法來建構優化系統。

本論文所提出方法的核心為，藉由前端精細分類器所產生的語意分割結果，能為循環收縮聚合技術(iteratively contractive merging)提供高階資訊，加上該技術於捕捉低階資訊優異的能力，得以重新產生更準確的語意分割。透過實驗結果分析，此方法之表現力可達到與目前最佳語意分割技術之相似水準，同時所產生之結果擁有更簡練的物件邊界。基於上述理由，循環收縮聚合技術有極大潛力得以取代條件隨機域。

The state-of-art models in semantic image segmentation usually contain a convolutional neural network (CNN) and a conditional random field (CRF) structure. As a dense predictor, CNN sometimes generates a prediction result with obvious boundary errors. On the other hand, as a refinement model, CRF could constantly improve the outcome by forcing the local labels to maintain the consistency. However, the boundary blurring effect of CRF has become an inherent defect. This fact suggests that CRF could be replaced by some other approaches. Our model uses the iteratively contractive merging (ICM) process to facilitate the semantic segmentation process given the prediction results from a dense predictor. The ICM process has been utilized for dealing with the unsupervised segmentation issues and this process can faithfully preserve the boundary information and maintain the consistency of local labels. Guided by the high-level information, we use the ICM process as a tool to grow image segments in a bottom-up way and to produce progressively more accurate outcomes in an iterative way. The achieved performance of the proposed approach is comparable to the state-of-the-art models while the proposed approach can generate more accurate boundaries. In conclusion, the experiment results show that the ICM process has the potential to be used as a substitute of the CRF model.