DEEP LEARNING BASED LINE SEGMENT DETECTION AND ITS APPLICATION USING TRANSFER LEARNING AND ITS TECHNIQUES

Abstract

While many artificial environments contain line segments, they are commonly employed in computer vision tasks. They add spatial and structural information to essentials. These traditional image edge-based line detectors, such as aperture-based methods, can detect lines relatively quickly and reasonably well. However, they tend to struggle in noisy or messy conditions. In contrast, learned line detectors can directly work on more complex images, but they are generally non-granular and are heavily reliant on wireframe lines. It describes an approach named Deep LSD (Deep Learning-based Line Segment Detector), which unifies the best of both worlds, yielding an accurate and robust line detector that can learn context-free without ground truth line annotations. A deep neural network passes through the whole image and gives a region of interest, which is then used to calculate the position and angle of that line.

Furthermore, we present an optimization strategy for enhancing the visible edge as the preferred location and point of perspective for better depth estimation accuracy. The system is evaluated on low-level line detection benchmarks and several challenging datasets for subsequent tasks. We propose a new line segment detection algorithm using LETR (Line Extraction with Transformer), which does not use any post-processing or heuristic techniques. While traditional edge or  connection point detection methods require post-processing and heuristics techniques to draw the final line segment, LETR adopts token query-based methods, self-identification mechanisms, and novel decoding methods to detect line segments directly. With its multi-measurement encoder-decoder architecture and a novel distance-based loss function, LETR improves line quality recognition. The self-listening process gradually takes place along the line through online learning. We outperform on benchmarks like Wireframe and York Urban in our tests.