A Lightweight Convolutional Neural Network for Salient Object Detection

Abstract

U-shape networks are widely used in salient object detection. Recently, CTDNet with a Comprehensive Triangular Decoder improved detection efficiency, which made some improvement with respect to the complexity and slow training of U-shape networks. However, CTDNet is still not lightweight enough, and the use of Global Average Pooling for top-level semantic features can lead to the loss of global structural information. This paper proposes Trilateral Enhanced Network (TENet), a faster salient detection model based on CTDNet, for industrial application. TENet uses MobileNetV3 as a backbone network so that TENet only needs 3.72M parameters, which lightweight the network consequently. TENet contains a feature fusion module called Channel Attraction Enhanced Feature Fusion Model, which integrates high-level semantics to improve accuracy. Additionally, Convolutional Block Feature Enhancement Module is proposed, which can further enhance accuracy. In comparison with CTDNet, TENet is a lightweight network with faster detection speed and more detection accuracy. TENet robustly detects defects in salient texture images, indicating insensitivity to texture interference. Experiments show TENet maintains strong performance on salient textures detection, demonstrating suitability for industrial optical inspection.

Keywords

• CTDNet
• industrial optical inspection
• MobileNetV3
• salient object detection
• TENet