RGB-D salient object detection with local feature and semantic segmentation

Abstract

Red, green, blue–depth (RGB-D) salient object detection (SOD) focuses on identifying visually prominent objects by simulating human visual perception. While existing RGB-D SOD methods have demonstrated results, there remain challenges in effectively leveraging extrinsic cues and enhancing feature representation. To address these limitations, novel RGB-D SOD model with local feature extraction and semantic segmentation (LFSS) is introduced, which is built on an encoder-decoder architecture. The encoder preprocesses the input images by merging RGB and depth data through a channel and spatial attention (CSA) module. A local feature extraction module further refines this fusion. The decoder consists of three key modules: i) the multi-feature extraction (MFE) module enhances base features through diverse convolutional operations; ii) the semantic segmentation enhancement (SSE) module optimizes features via spatial pyramid pooling and atrous convolution; and iii) the local/global agreement and edge detection (LGE) module that enables multi-level feature interaction and edge detection. These modules work sequentially to enhance and extract salient objects. LFSS is evaluated on six standard RGB-D SOD datasets (NJU2K, NLPR, STERE, LFSD, SSD, SIP) by four metrics, outperforming the comparison models with up to 1.2% F-measure improvement. LFSS is found to be a versatile model, offering valuable applications in engineering.