Article Preview
TopIntroduction
In recent years, rapid advancements in machine learning and deep learning have found extensive applications across various domains. For instance, supervised classification leveraging machine learning techniques has been explored (Salhi et al., 2021). Text analysis has notably benefited from deep learning methodologies (Singh & Sachan, 2021; Ismail et al., 2022; Gu et al., 2022), alongside sentiment analysis (Mohammed et al., 2022), industrial applications (Sharma et al., 2022), medical diagnostics (Xu et al., 2021), disease safety detection (Nguyen et al., 2021), and image enhancement tasks like defogging (Liu et al., 2022). The metaverse (Deveci et al., 2022) emerges as a groundbreaking platform for experimenting with autonomous driving, heavily reliant on deep learning for its core technology. Image segmentation and boundary detection are crucial in the field of computer vision, serving various fields such as autonomous driving assistance (Teichmann et al., 2018), simultaneous localization and mapping (SLAM) (Chen et al., 2021), point cloud segmentation (Wang et al., 2022), and medical imaging (Liu et al., 2022). Semantic segmentation involves assigning specific labels to individual pixels within objects, while boundary detection focuses on delineating object edges. However, prevalent neural network architectures like FCN (Long et al., 2015), ODE (Zhou et al., 2014), and UERF (Luo et al., 2016) face challenges in effectively capturing extensive pixel relationships as network depth increases, impeding accurate pixel classification. Moreover, deeper networks introduce noise and interference, further complicating the precise classification of minimal pixel clusters and leading to resolution loss and blurring during feature extraction downsampling. As network depth increases, external factors increasingly interfere with end-to-end segmentation. Predicted image outputs often contain unknown pixel classes, significantly impacting subject boundary segmentation quality. Accurate recognition of subject boundary pixels is crucial, especially for mobile robots operating in various environments. The main focus of boundary detection in the study aims to precisely locate subject boundary pixels, even in scenarios involving multiple object classes, where edge pixels belonging to different classes can lead to inadequate environmental understanding by mobile robots.
Figure 1. EPSSNet is compared with other lightweight models