Convolutional Encoder-Decoder Based Salient Object Detection

Saliency detection has achieved outstanding performance by utilizing deep encoder-decoder networks. Many of the existing saliency approaches, however, are unable to effectively process images with intricate backgrounds. In this study, saliency prediction is achieved using an autoencoder-based model. The new model utilizes the U-net design, which enables the exchange of information in both directions between the encoder and decoder. We assessed the suggested strategy by contrasting it with five cutting-edge saliency detection methods, utilizing the publically accessible MSRA10K dataset. The trials provide empirical evidence that the suggested strategy surpasses current advanced methodologies in AUC, F-measure, Recall, and Precision, while maintaining comparable performance for MAE. The effort employs transfer learning to streamline the training process.