SEC-UNET_SEMANTIC_EMBEDDING_AND_CONTOUR_ASSIST_UNET_FOR_BACTERIA_SEGMENTATION-AND-DETECTION
The number of bacterial types is a critical monitoring indicator for indoor air quality standards. It is a challenging task to cultivate and count colonies of bacteria which is expertise required and time-consuming. In this work, we investigate several U-Net improvement approaches. We are motivated by the assumption that contour information and semantic embedding branch can enhance U-Net's segmentation capacity for blurred and overlapping objects. Therefore, we propose Semantic Embedding and Contour Assist U-Net (SEC-U-Net) for direct bacteria segmentation and a shallow CNN for bacteria classification. This algorithm designed the detection of bacteria as a two-stage segmentation and classification task. Experimental results demonstrate the proposed method outperforms the state-of-the-art improved U-Net approaches on our bacteria dataset. Proposed SEC-U-NET+CNN based training presented over 91% and 85% precision rate for E.coli and S.aureus, respectively.
