LiteSalNet - Salient Object Detection

Abstract

Salient object detection in optical remote sensing images (ORSIs) presents unique challenges due to small object sizes, low contrast, and complex backgrounds. Existing methods often rely on computationally intensive architectures, limiting real-world applicability. LiteSalNet addresses these issues by introducing a lightweight MobileNetV2 backbone with only 3.90M parameters, combined with three core modules: the Adaptive Spatial Attention Module (ASAM), Dual-Scale Feature Enhancement Module (DSFEM), and Semantic Context Enhancement Module (SCEM). These modules work synergistically to refine multi-scale features through a Multi-Stream Progressively Decoding Framework (MSPDF), achieving state-of-the-art performance on ORSI datasets while maintaining computational efficiency.

Framework Architecture

Figure 1: Overall architecture of LiteSalNet. The network follows an encoder-decoder structure based on MobileNetV2, featuring three specialized modules: Adaptive Spatial Attention Module (ASAM), Dual-Scale Feature Enhancement Module (DSFEM), and Semantic Context Enhancement Module (SCEM). These modules refine multi-scale features, which are further processed through the Multi-Stream Progressively Decoding Framework (MSPDF). The MSPDF employs three parallel decoding streams for saliency prediction, edge detection, and skeleton estimation, enabling accurate, boundary-aware, and structure-consistent saliency detection.

Quantitative Results

Table 1: Quantitative comparison with SOTA methods on ORSI datasets. Top scores are highlighted in red (best), blue (second), and green (third).

PR Curve

Figure 2: Visual comparisons with 12 state-of-the-art ORSI-SOD models on various challenging ORSI scenarios. Please zoom-in for better visualization.

PR Curve

Figure 3: PR curves (left column) and F-measure curves (right column) on the ORSSD and EORSSD datasets.