WRGPruner: A new model pruning solution for tiny salient object detection

The model pruning is one of the predominant model compression tasks to decrease the demands in computing power and memory footprint. However, most existing pruning methods have overly broad application areas, which defects in a sub-optimal solution specifically to solve certain specified difficult problems in the tasks of salient object detection. In this paper, we propose a novel solution, dubbed as WRGPruner, based on the concept of salient energy level (SEL) for tiny salient object detection. The concept of SEL defines the level of assessing the distinguishing ability of parameters in the trained model between background and salient objects. To exploit the SEL, the WRGPruner is proposed, which considers three factors for model compression including the weight in the filter, the mathematical rank of the feature map matrix, and the gradient in the backward propagation. We mathematically prove the effectiveness of the WRGPruner for tiny salient objects. Besides, a tiny salient object dataset (TSOD) is constructed for evaluation. Extensive experiments show that WRGPruner reduces 60% of parameters with slight enhancement in terms of six accuracy metrics for VGG16 on TSOD. This demonstrates that the SEL is suitable for measure parameters and the effectiveness of WRGPruner.