CGAN: closure-guided attention network for salient object detection

In recent years, salient object detection (SOD) has achieved significant progress with the help of convolution neural network (CNN). Most of the state-of-the-art methods segment the salient object by either aggregating the multilevel features from the CNN module or introducing the refinement module along with the baseline network. However, these models suffer from simplicity bias, where neural networks converge to global minima using the simple feature and remain invariant to complex predictive features. Very few methods concentrate on the neurophysiological behaviour of visual attention. As per Gestalt psychology, humans tend to perceive the objects as a whole rather than focus on the discrete elements of that object. The law of Closure (closed contour) is one of the Gestalt axioms that states that if there is a discontinuity in the object’s contour, we perceive the object as continuous in a smooth pattern. This paper proposes a two-way learning network, where Closure-guided Attention Network (CGAN) and the Coarse Saliency Networks (CSN) jointly supervise the feature-channel to mitigate the simplicity bias. Furthermore, a channel-wise attention residual network is incorporated in the Closure Guided module to alleviate the scale-space problem and generate smooth object contour. Finally, the closure map from CGAN fused with the coarse saliency map of the Coarse Saliency Network generates a salient object. Experimental result on five benchmark datasets demonstrates the significant improvements in our approach over the state-of-the-art method.