Salient Object Detection via Hybrid Encoder and Feature Feedback Fusion Mechanism
Xueliang Li · 2023
Deep encoder-decoder networks have been adopted for saliency detection and achieved the state-of-the-art performance. However, most existing saliency methods usually fail to process images with complex background and small objects. In this paper, we propose and apply a new method for saliency detection using hybrid encoder and feature feedback fusion mechanism. The proposed network architecture is composed of hybrid encoder based feature extraction module, feature feedback fusion mechanism based decoder module and multi observation point supervised learning module. In the feature extraction module, CNN encoder and transformer encoder are used to extract the local and global features of the image respectively. The decoder module optimizes the salient features of each layer by effectively fusing the visual features of the two layers or three layers. The multi observation point supervised learning module sets multiple observation points at each layer of the decoder, and each observation point considers the binary cross entropy loss and intersection over union loss. A large number of experimental results on five public data sets show that the method proposed in this paper is significantly better than other methods in different evaluation metrics.