Abstract
This research introduces "flyweight" Salient Object Detection (SOD) networks designed for extreme efficiency and minimal data requirements. By utilizing a backpropagation-free encoder (FLIM) and novel adaptive decoders that estimate weights per image via heuristics, these models can be configured with as few as three or four images. The approach achieves state-of-the-art performance on specialized biological and medical imaging tasks, significantly reducing the human annotation burden.
Introduction
Salient Object Detection (SOD) is critical for identifying prominent objects and generating precise segmentation masks. Despite the success of deep learning, standard models require massive computational resources and thousands of manually labeled images. This is particularly problematic in niche domains like microscopy or medical imaging, where expert annotation is scarce. This paper proposes a "flyweight" architecture that is hundreds of times more efficient than current lightweight models and can be trained in minutes without backpropagation.
Methodology
FLIM Encoder (Feature Learning from Image Markers)
The core of the architecture is the FLIM encoder, which avoids gradient descent. Instead, it relies on user-provided markers (dots or lines) to identify foreground and background regions in a small set of training images. The network applies k-means clustering on patches surrounding these markers to define convolutional kernels.
Adaptive Decoders
The innovation lies in the adaptive decoders, which use heuristic functions to estimate weights for each input image at inference time. The authors introduce Probability-based (pb) and Mean-based (mb) decoders. Unlike previous iterations that used global weights, these new decoders estimate unique weights for each neuron per pixel, providing granular control over the saliency map.
Graph-based Delineation
To ensure the saliency maps align with physical object boundaries, a "Dynamic Trees" algorithm is employed as a post-processing step to refine the final segmentation.
Experimental Evaluation
Datasets
The models were tested on two specialized datasets:
- Parasites: 1,219 RGB images of Schistosoma Mansoni eggs. The challenge involves distinguishing eggs from similar-looking debris.
- BraTS (Brain Tumor Segmentation): 3,743 2D grayscale MRI slices featuring high variance in tumor characteristics.
Key Results
The FLIM-based models demonstrated significant advantages:
- Efficiency: Configurations with as few as 50,000 parameters outperformed models like SAMNet (~1.33M parameters).
- Data Scarcity: Training on only 3-4 representative images yielded results competitive with or superior to state-of-the-art models trained on entire datasets.
- Robustness: Qualitative analysis showed fewer convergence errors compared to traditional models when data is limited or imbalanced.
Discussion and Conclusion
The research highlights that well-designed heuristics can effectively replace thousands of backpropagation iterations in specific domains. While deeper encoders reduce false positives, they can sacrifice boundary precision, necessitating separate delineation steps. Future work will focus on automating the selection of representative training images and further balancing decoder complexity with the flyweight philosophy.
Summary updated