170 J. D. Eckels et al. 18.2.2 Image Segmentation Image segmentation, or semantic segmentation, is an important task in the field of computer vision where regions or pixels in an image are classified and color-coded by class or object type. Image segmentation has seen significant development recently in several different fields, including autonomous driving, robotic navigation, remote sensing, and medical imaging [26]. Deep neural networks have shown promising results when applied to image segmentation tasks, including the use of conditional random fields (CRFs) [27, 28], feedback CNNs [29], feedforward labeling of superpixels [30], region-based CNNs (R-CNNs) [31], flexible segmentation graphs (FSGs) [32], fully convolutional networks (FCNs) [33], and up-sampling and deconvolutional layers with multi-path fusion [34–36]. Ronneberger et al. [37] proposed a U-Net style architecture for image segmentation with symmetry between the down-sampling (encoder) and up-sampling (decoder) paths of the neural network, as shown in Fig. 18.1. Feature maps in the encoder are carried over to the decoder by the horizontal paths in Fig. 18.1 for improved pixel-wise prediction and resolution in the output segmentation map. The U-Net model is a type of FCN, where all model parameters in both the encoder and decoder can be optimized by training the network on a labeled image segmentation dataset. An altered version of the U-Net style architecture was used in the current study. Further detail on image segmentation and object detection can be found in the surveys by Lateef and Ruichek [26] and Liu et al. [38]. 18.3 Methodology 18.3.1 Project Overview Figure 18.2 provides a high-level conceptual map detailing the flow of data between each working subprocess and software component in this study. Fig. 18.1 U-Net style CNN architecture for image segmentation [37]
RkJQdWJsaXNoZXIy MTMzNzEzMQ==