An Attention Based Wavelet Convolutional Model for Visual Saliency Detection

The emergence of deep neural architectures greatly enhanced the accuracy of salient region detection algorithms that plays a vital role in computer vision applications. However, the accurate extraction of regions with fine boundaries still remains as a challenge. In this work, an attention based Wavelet Convolutional Neural Network (WCNN) is implemented that efficiently extracts the spatial, spectral and semantic features of the image in multiple resolution and it turns out to be suitable for locating the visually salient regions. Further enhancement of the fine boundaries of the predicted map is made possible by the inclusion of a combinational loss function of balanced cross entropy loss, SSIM loss and edge loss. The effectiveness of the method is evaluated using three benchmark datasets and the results shows better performance achieving a minimum Mean Absolute Error (MAE) of 0.032 and maximum F-measure of 0.938.