Adaptive Decoders for FLIM-Based Salient Object Detection Networks

Salient Object Detection (SOD) methods based on deep learning have succeeded, usually at the price of abundantly annotated data and intensive computational resources. Such limitations have motivated the development of lightweight models, but they are still pre-trained on large datasets, and their adaptation under labeled data scarcity is challenging. In this context, Feature Learning from Image Markers (FLIM) is a methodology under investigation to create convolutional encoders with minimal user effort in data annotation. Flyweight networks based on a FLIM encoder followed by an adaptive decoder, which is a point-wise convolution with adaptive weights for each image followed by activation, achieved state-of-the-art results for SOD recently. In this work, we propose four strategies for computing adaptive weights based on (i) channel-tri-state detection, (ii) labeled markers, (iii) channel attention, and (iv) a hybrid solution using the tri-state and labeled-marker decoders. An assessment on two medical datasets between FLIM-based SOD networks with the proposed adaptive decoders, three state-of-the-art lightweight models and a U-shaped network with a FLIM encoder has shown that the results favor FLIM networks, with the hybrid solution being the most promising option.