SANet: scale-adaptive network for lightweight salient object detection

Abstract

Salient object detection (SOD) is widely used in transportation such as road damage detection, assisted driving, etc. However, heavyweight SOD methods are difficult to apply in scenarios with low computing power due to their huge amount of computation and parameters. The detection accuracy of most lightweight SOD methods is difficult to meet application requirements. We propose a novel lightweight scale-adaptive network to achieve a trade-off between lightweight restriction and detection performance. We first propose the scale-adaptive feature extraction (SAFE) module, which mainly consists of two parts: multi-scale feature interaction, which can extract features of different scales and enhance the representation ability of the network; and dynamic selection, which can adaptively assign different weights to features of varying scales according to their contribution through the input image. Then, based on the SAFE module, a lightweight and adaptive backbone network is designed, and scale-adaptive network is implemented in combination with the multi-scale feature aggregation (MFA) module. We evaluate the model quantitatively and qualitatively on six public datasets and compare it with typical heavyweight and lightweight methods. With only 2.29 M parameters, it can achieve a prediction speed of 62 fps on a GTX 3090 GPU, far exceeding other models, and real-time performance is guaranteed. The model performance reaches that of general heavyweight methods and exceeds state-of-the-art lightweight methods.

Keywords

Salient object detection / lightweight SOD / model lightweighting / multi-scale learning