8 October 2024 in IEEE Access
We present a salient object detection method in this paper. It uses a lightweight network based on multi-scale information aggregation (MIA) module and bidirectional cross-stage feature fusion (BCF2) module to speed up the reasoning process. Besides, it designs a stepwise coupling optimization strategy based on positive-incentive pseudo label to alleviate the miss detection rate for small objects and the false detection rate for big objects. It isworth noting that the positive-incentive pseudo label generation mechanism can increase the ratio of salient pixels in small objects to alleviate the severe category imbalance problem, while the stepwise coupling optimization strategy can avoid the misguidance of pseudo label to big objects. In addition, the experimental results demonstrate that our method is fairly efficient, its frame rate in reasoning phase reaches 275.6 FPS on NVIDIA RTX3080 GPU when the input resolution is set to 336 × 336.
17 June 2025 in IEEE Access
Recently, salient object detection (SOD) methods based on deep learning have achieved significant progress.The existing models often require a considerable amount of memory and time to make precise inferences, SOD is an image preprocessing task, it is obviously unreasonable to consume a large amount of cost on SOD, which makes previous methods difficult to apply in practical scenarios. In this paper, we aim to find a balance between performance and efficiency by designing a lightweight network based on feature pyramid network (FPN), namely MII-FPN, to alleviate this problem. MII-FPN mainly contains a simple multi-scale convolution module and an efficient feature fusion module, which are used to extract and integrate the context information, respectively. Testing on a commonly used six datasets shows that MII-FPN achieves highly competitive accuracy compared with state-of-the-art SOD methods.Besides, the proposed MII-FPN is fairly efficient, which has a frame rate of 299.3FPS on NVIDIA RTX3080 GPU for 336 × 336 input with only 1.12M parameters. The code is available at https://github.com/XUPTfangjie/MIIFPN.