Abstract: Salient instance segmentation is to segment the most noticeable instance object in the image. However, there remain some problems in the existing methods of salient instance segmentation. For example, the small salient instances are difficult to be detected and segmented, and the segmentation accuracy is insufficient for large salient instances. Therefore, to solve these two problems, a new salient instance segmentation model, namely the attention residual multi-scale feature enhancement network (ARMFE), has been proposed. ARMFE includes two modules, i.e. the attention residual network module and the multi-scale feature enhancement module. The attention residual network module combines the residual network with the spatial attention sub-module and the channel attention sub-module to enhance the features. The multi-scale feature enhancement module can further enhance the information fusion for features with large scale span based on the feature pyramid. Therefore, the proposed ARMFE model can make full use of the complementary information of multi-scales features by attention residual multi-scale feature enhancement, and then simultaneously improve the accuracy of detecting and segmenting large instance objects and small instance objects. The proposed ARMFE model has been tested on the salient instance segmentation dataset Salient Instance Saliency-1K (SIS-1K), and the segmentation accuracy and speed have been improved. This indicates that our proposed model outperforms other existing salient instance segmentation algorithms, such as MSRNet and S4Net. 

Key words: salient instance segmentation, attention mechanism, residual network, multi-scale, feature enhancement