Researchers have developed a new approach to help computers better recognize textures in images—an important task in fields ranging from industrial quality control to medical diagnosis. Their method, called FLEX-LBP, refines a classic technique known as Local Binary Patterns (LBP) by teaching the system to automatically adjust its settings for each dataset, rather than relying on fixed, hand-tuned parameters. This leads to more accurate texture classification while keeping the computational cost low and the results easy to interpret.
Key Takeaways
- FLEX-LBP learns 90 parameters—including sampling radius, angles, color transformations, and thresholds—using backpropagation, a standard AI training technique.
- The method produces exact binary codes during processing, maintaining the interpretability and efficiency of traditional LBP descriptors.
- On six different color texture datasets, FLEX-LBP improved nearest-neighbor classification accuracy by an average of 3.55 percentage points over fixed LBP, and outperformed several classical and neural network-based texture descriptors.
- On a medical imaging benchmark for colorectal cancer histology, FLEX-LBP achieved 93.8% accuracy, surpassing previous best classical methods.
Local Binary Patterns (LBP) are a popular way for computers to describe textures in images. They work by comparing each pixel with its neighbors and encoding these comparisons into a simple binary pattern—a string of zeros and ones—that captures the local texture. This method is fast, easy to interpret, and effective, which is why it’s used in many applications such as detecting defects on manufacturing lines or analyzing tissue samples in healthcare.
However, traditional LBP requires manually choosing several parameters like the radius around each pixel to sample, the angles at which comparisons are made, the color channels to use, and the threshold values for deciding whether neighbors are brighter or darker. These fixed choices may not work well across different datasets or imaging conditions.
The FLEX-LBP approach tackles this limitation by allowing the system to learn these parameters automatically during training. Using a process called backpropagation—commonly employed in training neural networks—the method adjusts the parameters to optimize texture classification accuracy. A key innovation is that FLEX-LBP still produces exact binary codes during the forward pass (the step where the image is analyzed), preserving the simplicity and interpretability of the original LBP. To handle the challenge of learning from binary outputs, the researchers use a “straight-through estimator” during backpropagation, which approximates how to update parameters despite the discrete nature of the binary codes.
In testing, FLEX-LBP was evaluated on six standard color texture benchmarks, covering a variety of image types and conditions. The researchers found that learning parameters related to geometry and thresholds, while keeping the color transformation fixed, consistently improved accuracy. When combined with a linear discriminant analysis step (a simple classification method), performance improved further. The fully learned version of FLEX-LBP, which adjusts all parameters, achieved the highest average accuracy. Notably, FLEX-LBP outperformed several classical LBP variants, a neural LBP approach, and even features extracted from a large neural network pretrained on ImageNet, despite FLEX-LBP’s much smaller size and computational simplicity.
One striking result came from the Kather colorectal cancer histology dataset, where FLEX-LBP reached 93.8% accuracy—beating the best classical descriptor by a significant margin. The study also examined robustness to noise and lighting changes, showing that certain preprocessing steps like median sampling or gray-world normalization can further stabilize performance.
This research highlights the value of combining classical, interpretable image descriptors with modern learning techniques. By automatically tuning its parameters, FLEX-LBP adapts to different datasets without losing the efficiency and clarity that make LBP popular. The authors have made their code available as a convenient PyTorch layer, enabling easy adoption by the research community and industry practitioners.
Looking ahead, FLEX-LBP’s balance of accuracy, speed, and interpretability could make it a useful tool in applications where understanding texture is critical but computational resources are limited. This includes real-time industrial inspection, medical image analysis on standard hardware, and other domains where deploying large neural networks is impractical. Further research may explore combining FLEX-LBP with other methods or extending it to video and 3D texture analysis.
Based on research published on arXiv by Mohamed Alimoussa, Amal El Kaid.