Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FLIM-based Salient Object Detection Networks with Adaptive Decoders

Gilson Junior Soares, Matheus Abrantes Cerqueira, Jancarlo F. Gomes, Laurent Najman, Silvio Jamil F. Guimarães, Alexandre Xavier Falcão

TL;DR

FLIM-based flyweight SOD networks address salient object detection under limited labeled data and constrained computation by pairing a marker-trained FLIM encoder with adaptive decoders that estimate weights per input. The approach introduces five adaptive decoders (including two new per-pixel variants) and trains encoders from only a few representative images without backpropagation, evaluated on parasite egg and BraTS brain tumor datasets. Results show the FLIM models with adaptive decoders outperform fixed-weight decoders and several lightweight baselines, demonstrating strong data efficiency and robustness across tasks. The work highlights practical implications for resource-constrained settings and suggests future directions in broader per-pixel adaptation and delineation strategies.

Abstract

Salient Object Detection (SOD) methods can locate objects that stand out in an image, assign higher values to their pixels in a saliency map, and binarize the map outputting a predicted segmentation mask. A recent tendency is to investigate pre-trained lightweight models rather than deep neural networks in SOD tasks, coping with applications under limited computational resources. In this context, we have investigated lightweight networks using a methodology named Feature Learning from Image Markers (FLIM), which assumes that the encoder's kernels can be estimated from marker pixels on discriminative regions of a few representative images. This work proposes flyweight networks, hundreds of times lighter than lightweight models, for SOD by combining a FLIM encoder with an adaptive decoder, whose weights are estimated for each input image by a given heuristic function. Such FLIM networks are trained from three to four representative images only and without backpropagation, making the models suitable for applications under labeled data constraints as well. We study five adaptive decoders; two of them are introduced here. Differently from the previous ones that rely on one neuron per pixel with shared weights, the heuristic functions of the new adaptive decoders estimate the weights of each neuron per pixel. We compare FLIM models with adaptive decoders for two challenging SOD tasks with three lightweight networks from the state-of-the-art, two FLIM networks with decoders trained by backpropagation, and one FLIM network whose labeled markers define the decoder's weights. The experiments demonstrate the advantages of the proposed networks over the baselines, revealing the importance of further investigating such methods in new applications.

FLIM-based Salient Object Detection Networks with Adaptive Decoders

TL;DR

FLIM-based flyweight SOD networks address salient object detection under limited labeled data and constrained computation by pairing a marker-trained FLIM encoder with adaptive decoders that estimate weights per input. The approach introduces five adaptive decoders (including two new per-pixel variants) and trains encoders from only a few representative images without backpropagation, evaluated on parasite egg and BraTS brain tumor datasets. Results show the FLIM models with adaptive decoders outperform fixed-weight decoders and several lightweight baselines, demonstrating strong data efficiency and robustness across tasks. The work highlights practical implications for resource-constrained settings and suggests future directions in broader per-pixel adaptation and delineation strategies.

Abstract

Salient Object Detection (SOD) methods can locate objects that stand out in an image, assign higher values to their pixels in a saliency map, and binarize the map outputting a predicted segmentation mask. A recent tendency is to investigate pre-trained lightweight models rather than deep neural networks in SOD tasks, coping with applications under limited computational resources. In this context, we have investigated lightweight networks using a methodology named Feature Learning from Image Markers (FLIM), which assumes that the encoder's kernels can be estimated from marker pixels on discriminative regions of a few representative images. This work proposes flyweight networks, hundreds of times lighter than lightweight models, for SOD by combining a FLIM encoder with an adaptive decoder, whose weights are estimated for each input image by a given heuristic function. Such FLIM networks are trained from three to four representative images only and without backpropagation, making the models suitable for applications under labeled data constraints as well. We study five adaptive decoders; two of them are introduced here. Differently from the previous ones that rely on one neuron per pixel with shared weights, the heuristic functions of the new adaptive decoders estimate the weights of each neuron per pixel. We compare FLIM models with adaptive decoders for two challenging SOD tasks with three lightweight networks from the state-of-the-art, two FLIM networks with decoders trained by backpropagation, and one FLIM network whose labeled markers define the decoder's weights. The experiments demonstrate the advantages of the proposed networks over the baselines, revealing the importance of further investigating such methods in new applications.
Paper Structure (32 sections, 15 equations, 13 figures, 5 tables, 1 algorithm)

This paper contains 32 sections, 15 equations, 13 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. FLIM encoders
  4. Images and patches
  5. Sets of marker pixels
  6. Convolutional block and FLIM encoder
  7. Marker-based normalization
  8. Convolution, activation, and pooling
  9. Adaptive Decoders
  10. Tri-state adaptive decoder ()
  11. Attention-based adaptive decoder ()
  12. Label-based tri-state adaptive decoder ()
  13. Probability-based adaptive decoder ()
  14. Mean-based adaptive decoder ()
  15. Experimental setup
  16. ...and 17 more sections

Figures (13)

  • Figure 1: Marker drawing for kernel estimation: (a) A training image with foreground (red) and background (white) markers; (b) background activation channel from a background kernel; (c) foreground activation channel from an object kernel; (d) resulting saliency map.
  • Figure 2: A FLIM-based SOD model generalizing to new images. (a) A test image where the red arrow indicates the object (a parasite egg) and (b) the resulting saliency map.
  • Figure 3: Adaptive decoders. (a) A point-wise convolution followed by activation, whose weights are estimated in ${-1,+1}$ by a heuristic function. (b) It defines one neuron per pixel, and all neurons share the same weights.
  • Figure 4: Saliencies generated by an adaptive decoder for different blocks of an architecture. The ground-truth's border is presented in magenta.
  • Figure 5: Pipeline used for the experiments. (a) The dataset $\mathcal{Z}$ is randomly divided into sets $\mathcal{Z}_1$ and $\mathcal{Z}_2$. (b) A few representative images are selected for the set $\mathcal{T}$, according to the model's performance on the validation set $\mathcal{Z}_1\backslash \mathcal{T}$. (c) Once $\mathcal{T}$ is fixed, the best network architecture is found on the validation set. (d) The pre-trained FLIM encoder with each adaptive decoder is tested on $\mathcal{Z}_2$. (e) Each fixed-weight decoder is trained on $\mathcal{T}$, using the pre-trained FLIM encoder as the fixed backbone; the best models are found on $\mathcal{Z}_1\backslash \mathcal{T}$ and tested on $\mathcal{Z}_2$. (f) Each pre-trained lightweight model is fine-tuned on $\mathcal{T}$; the best models are found on $\mathcal{Z}_1\backslash \mathcal{T}$ and tested on $\mathcal{Z}_2$.
  • ...and 8 more figures