Enhancing Salient Object Detection Through VGG16–VGG19 Architectures

Abstract

Convolutional Neural Networks (CNNs)-based deep learning techniques have significantly advanced Computer Vision (CV), particularly in Salient Object Detection (SOD). This study investigates two CNN architectures; VGG16 and VGG19 for SOD implementation using the MSRA-10K dataset. The models were trained on 1,160 and 2,700 images to assess the impact of dataset size on detection accuracy. Furthermore, two feature fusion strategies were evaluated: the averaged feature fusion (AvgVGG1619) and the maximum feature fusion (MaxVGG1619). The results indicate that VGG16 generally achieves superior performance, whereas VGG19 performs better on images lacking salient regions. Increasing the dataset size improved the evaluation metrics and significantly reduced the Mean Absolute Error (MAE). Among all models, MaxVGG1619 demonstrated the highest overall accuracy, even when trained on fewer images. Overall, these findings demonstrated the benefits of model fusion and dataset scaling in enhancing CNN-based SOD performance, and demonstrate the effectiveness of combining VGG models, particularly through using maximum feature selection, for robust and accurate salient object detection.