ACPSOD-Net: a deep atrous convolution pooling based network for salient object detection
Bhagyashree V. Lad, Mohammad Farukh Hashmi, Avinash G. Keskar · 2024
Systems incorporating a convolutional neural network (CNN) architecture are frequently utilized by many salient object detection (SOD) applications. When identifying the salient objects, detecting objects in diverse imaging scenarios is vital. To efficiently extract features and fuse them effectively, current SOD techniques generally use CNNs. Numerous strategies have been studies to further improve CNN representation in light of recent research that shows how well CNNs handle the edges and texture of images. This chapter introduces an original technique for identifying salient objects based on atrous convolution-based pooling and feature fusion networks to effectively integrate various features of objects. The proposed architecture is computationally very light and requires fewer learnable parameters. The chapter proposes a method based on an encoder–decoder network to detect salient objects. The encoders belong to the five ResNet-34 architecture's encoder blocks, and feature maps obtained by the encoders are passed through the atrous convolution-based pooling (ACP) block. The various features at the encoder step are integrated using a fusion network to obtain better saliency results. The fused feature maps are passed through five decoders to obtain the resultant saliency map. The outcomes of this study are assessed upon three well-known salient object detection datasets which illustrate the potential of the suggested system to accurately detect and localize the salient objects in different imaging scenarios.