Saliency detection with deep learning in 2 dimension images

The detection of objects in image data has become an important topic with the advancement of technology. Salient object detection aims to identify the most attention-grabbing object in an image. In complex backgrounds or when objects are overlapping, clearly detecting object boundaries becomes a challenging task. In the literature, various methods have been proposed that rely on object features such as texture, brightness levels, and color to detect salient objects. However, while these features can generally classify objects, they have not been successful in images with complex backgrounds. Recently, methods that utilize different preprocessing techniques and neural network architectures with various backbone networks have been widely used in saliency detection. However, objects' overall structure and fine details differ. Therefore, approaches combining neural networks with attention mechanisms have been developed. Inspired by natural language processing, transformer architectures have recently begun to be applied to image data. In this study, saliency detection was performed using the DUTS and ECSSD datasets. First, a Convolutional Neural Network (CNN) was used, followed by a transformer-based saliency detection approach. To enhance the performance of transformer architectures, data augmentation techniques such as noise injection, rotation, and blurring were applied to increase dataset diversity. In this way, the study also investigated the impact of preprocessing methods on model performance. For saliency detection, a segmentation method was used. Spatial attention helps the model focus on important regions in an image by determining which areas are most significant. Channel-wise attention evaluates the importance of each feature map and helps the model focus on more meaningful channels. The best results were obtained on the DUTS dataset using the "Segment Anything Model UNet - 2.1," achieving 0.019 Mean Absolute Error (MAE), 0.961 Enhanced Alignment Measure (E-measure), and 0.936 Structural Similarity Index (SSIM).