SSDBN: A Single-Side Dual-Branch Network with Encoder–Decoder for Building Extraction

Abstract

In the field of building detection research, an accurate, state-of-the-art semantic segmentation model must be constructed to classify each pixel of the image, which has an important reference value for the statistical work of a building area. Recent research efforts have been devoted to semantic segmentation using deep learning approaches, which can be further divided into two aspects. In this paper, we propose a single-side dual-branch network (SSDBN) based on an encoder–decoder structure, where an improved Res2Net model is used at the encoder stage to extract the basic feature information of prepared images while a dual-branch module is deployed at the decoder stage. An intermediate framework was designed using a new feature information fusion methods to capture more semantic information in a small area. The dual-branch decoding module contains a deconvolution branch and a feature enhancement branch, which are responsible for capturing multi-scale information and enhancing high-level semantic details, respectively. All experiments were conducted using the Massachusetts Buildings Dataset and WHU Satellite Dataset I (global cities). The proposed model showed better performance than other recent approaches, achieving an F1-score of 87.69% and an IoU of 75.83% with a low network size volume (5.11 M), internal parameters (19.8 MB), and GFLOPs (22.54), on the Massachusetts Buildings Dataset.

An Attentive Hough Transform Module for Building Extraction From High Resolution Aerial Imagery

“…In the assessment of generalization performance, we utilized the WHU Satellite Dataset I. The results presented in Table Furthermore, a direct comparison to SSDBN [36] in Table 5 reveals our model's superior performance, particularly in IoU and F1-score metrics. Notably, our model exhibited a significant 15.9 % increase in F1-score compared to SSDBN [36], underscoring positive implications for its generalization ability.…”

“…Additionally, with the emergence and advancement of attention mechanisms such as spatial attention [23], self-attention [24], squeeze and excitation networks [25] and Convolutional Block Attention Module (CBAM) [26], significant progress has been achieved in building extraction. Various studies have leveraged these attention mechanisms to boost network capabilities, contributing unique insights into the field [19], [27]- [36].…”

“…They then employ squeeze-and-excitation attention and lightweight residual blocks, to refine and reconstruct building features in the decoding stage. In [36], the authors introduce the SSDBN model, incorporating an enhanced Res2Net encoder and a dual-branch decoder with CBAM. This design emphasizes the model's proficiency in capturing both global and local context details.…”

CFNet: An Eigenvalue Preserved Approach to Multiscale Building Segmentation in High-Resolution Remote Sensing Images

“…Swish was proposed by Google researchers and is computationally as efficient as ReLU and shows better performance than other functions for deeper models. Swish is shown in Equation (2).…”

“…S EMANTIC segmentation is a fundamental research problem in today's computer vision research community, which aims to segment images into several predefined and semantically labeled coherent parts [1], [2]. This is a specific application of artificial intelligence in a realistic sense, and is used in radar [3], [4], remote sensing [5], and electricity [6].…”