Flyweight FLIM Networks for Salient Object Detection in Biomedical Images

TL;DR

Biomed SOD faces data scarcity and resource constraints; the paper introduces Flyweight FLIM networks that learn encoder kernels from a handful of markers without backpropagation, extended with dilated-separable layers and a dynamic adaptive decoder. It adds a network simplification technique to prune redundant kernels, producing models with far fewer parameters and FLOPs while maintaining competitive accuracy against heavyweight SOTA on two challenging biomedical datasets. The method leverages information redundancy in biomedical images and marker-guided patch clustering to achieve efficient, specialised models suitable for deployment on low-power devices. Experimental results show substantial efficiency gains and strong performance relative to lightweight baselines, with competitive results against heavyweight models in several settings. These findings highlight the practical potential of FLIM-based flyweight networks for data-limited, resource-constrained biomedical imaging tasks.

Abstract

Salient Object Detection (SOD) with deep learning often requires substantial computational resources and large annotated datasets, making it impractical for resource-constrained applications. Lightweight models address computational demands but typically strive in complex and scarce labeled-data scenarios. Feature Learning from Image Markers (FLIM) learns an encoder's convolutional kernels among image patches extracted from discriminative regions marked on a few representative images, dismissing large annotated datasets, pretraining, and backpropagation. Such a methodology exploits information redundancy commonly found in biomedical image applications. This study presents methods to learn dilated-separable convolutional kernels and multi-dilation layers without backpropagation for FLIM networks. It also proposes a novel network simplification method to reduce kernel redundancy and encoder size. By combining a FLIM encoder with an adaptive decoder, a concept recently introduced to estimate a pointwise convolution per image, this study presents very efficient (named flyweight) SOD models for biomedical images. Experimental results in challenging datasets demonstrate superior efficiency and effectiveness to lightweight models. By requiring significantly fewer parameters and floating-point operations, the results show competitive effectiveness to heavyweight models. These advances highlight the potential of FLIM networks for data-limited and resource-constrained applications with information redundancy.

Figures (17)

• Figure 1: (a) Original Image with background and object components. User-drawn markers are shown in cyan (object) and red (background); (b) Foreground activation channel with the object (yellow arrow) and some false positives (pink arrows) activated; (c) Background activation channel, in which the object (yellow arrow) is not activated; (d) Resulting saliency map from an adaptive decoder.
• Figure 2: Comparison of model's size and performance among FLIM and baselines from the state-of-the-art for SOD on the S. mansoni eggs dataset. The size of the circles represents the number of Giga Floating Point Operations (GFLOPS).
• Figure 3: Steps for learning a flyweight FLIM layer are shown. Green arrows illustrate the output of each layer, while black arrows indicate possible data flows, with dotted lines representing paths used for training a simplified, non-separable CNN.
• Figure 4: Diagram of kernel factorization into depthwise separable ones.
• Figure 5: Diagram of kernel factorization into pointwise separable ones.