PaperlayerView exact revision on arXiv

Source-linked AI summary

BoWFire: Detection of Fire in Still Images by Integrating Pixel Color and Texture Analysis

Daniel Y. T. Chino, Letricia P. S. Avalhais, Jose F. Rodrigues, Agma J. M. Traina

arXiv:1506.03495v1cs.CV

TL;DR

Fire-detection systems must filter large volumes of crisis imagery, while video-oriented color models can generate false positives on still images and be difficult to tune. BoWFire combines color classification with texture classification on superpixel regions and uses fewer parameters than prior methods. Results show reduced false positives with precision compatible with state-of-the-art methods.

  • Problem

    Still-image fire detection needs to filter irrelevant crowdsourced content, but video-oriented color models can produce high false-positive rates and difficult-to-tune parameters.

  • Method

    BoWFire combines pixel-color classification with texture classification on superpixel regions and merges detections retained by both classifiers.

  • Results

    BoWFire reduces false positives while maintaining performance similar to state-of-the-art methods, with texture addition reducing false-positive rates by up to 80%.

  • Takeaways & Limitations

    Combining color and texture information can discard false positives while preserving fire-detection performance in still images.

Abstract

from arXiv · show

Emergency events involving fire are potentially harmful, demanding a fast and precise decision making. The use of crowdsourcing image and videos on crisis management systems can aid in these situations by providing more information than verbal/textual descriptions. Due to the usual high volume of data, automatic solutions need to discard non-relevant content without losing relevant information. There are several methods for fire detection on video using color-based models. However, they are not adequate for still image processing, because they can suffer on high false-positive results. These methods also suffer from parameters with little physical meaning, which makes fine tuning a difficult task. In this context, we propose a novel fire detection method for still images that uses classification based on color features combined with texture classification on superpixel regions. Our method uses a reduced number of parameters if compared to previous works, easing the process of fine tuning the method. Results show the effectiveness of our method of reducing false-positives while its precision remains compatible with the state-of-the-art methods.

I. INTRODUCTION

Fire emergencies motivate automated analysis of crowdsourced images, but video-oriented color models can produce false positives on still images. BoWFire addresses this by combining pixel-color and texture classification with fewer tuning parameters.

  • Fire emergencies can cause economic, environmental, and human harm, motivating automatic fire detection systems.
  • Crowdsourced images and videos can provide crisis information, but their volume requires automated filtering of irrelevant content.
  • Video fire-detection methods combine color-based frame analysis with temporal features, but still-image processing lacks temporal information.
  • Color-only models can mistake highly illuminated or reddish-yellowish objects for fire, increasing false positives.
  • BoWFire combines pixel-color classification and texture classification to detect fire in still images using visual information alone.
  • The method uses fewer parameters than former works, simplifying fine tuning while targeting fewer false positives.

II. RELATED WORKS

Prior fire-detection methods rely mainly on color, clustering, geometric features, or reference comparisons. BoWFire extends these approaches with texture information and fewer, more intuitive tuning parameters.

  • Chen’s method uses three RGB- and HSI-based rules with two user-set threshold parameters for fire-pixel detection.
  • Celik’s method defines five mathematical rules in YCbCr space to improve discrimination of fire pixels.
  • Rossi’s method combines YUV clustering with a 3D-Gaussian pixel classifier, assuming fire is recorded in a controlled environment.
  • Rudz’s method clusters YCbCr pixels and removes false positives by comparing small and large regions with reference statistics.
  • BoWFire adds texture beyond color and uses fewer parameters whose physical meaning makes fine tuning more intuitive.

III. BASIC CONCEPTS AND NOTATIONS

The paper represents images as sets of RGB pixels and describes feature extraction and supervised classification in vector and label spaces.

  • An image is represented as a set of pixels, with each pixel storing red, green, and blue channel intensities.
  • A feature extractor maps an image to a d-dimensional feature vector.
  • A supervised classifier learns from labeled training tuples and predicts a label for a new feature vector.

A. Feature Extraction

Image features describe visual properties globally or locally, including color, shape, and texture. The section highlights color descriptors and Local Binary Pattern texture features.

  • Extracted features are numerical measurements describing visual properties and relationships among pixels globally or in local regions.
  • Color features are computationally inexpensive and include color layout and histogram-derived color moments.
  • Local Binary Pattern extracts texture information from a pixel’s neighborhood by comparing neighboring intensities with the central pixel.
  • Rotation-invariant LBP shifts each binary code to its minimum value, while uniform LBP groups patterns with at most two bitwise transitions.

B. Superpixel Generation

Superpixels partition an image into contiguous, homogeneous regions that preserve boundaries while reducing redundancy and the complexity of later processing. BoWFire uses SLIC, a K-Means adaptation guided by Lab or YCbCr color values and geometric position.

  • B. Superpixel Generation: Superpixels group pixels into continuous regions with similar homogeneity, providing building blocks for image-segmentation algorithms.They capture image redundancy and reduce the complexity of subsequent image-processing tasks.
  • B. Superpixel Generation: A superpixel-generation algorithm maps image I and Ksp to Ksp image partitions, while aiming to adhere to boundaries and improve segmentation.
  • B. Superpixel Generation: SLIC adapts K-Means for superpixel generation by combining pixel color values with geometric position in a distance function.Users specify the number of superpixels Ksp and compactness m; centroids are placed on a regular grid away from edge pixels.

IV. OUR PROPOSAL

BoWFire detects fire in still images by combining pixel-color classification with texture classification over superpixel regions, then merging their outputs. The design targets fire-like regions that color alone may misclassify while limiting parameterization and retaining modularity.

  • IV. OUR PROPOSAL: BoWFire combines pixel-color classification and texture classification to distinguish actual fire from reddish or yellowish fire-like regions.The method is intended for automated fire-scene detection in emergency-situation images without temporal information.
  • IV. OUR PROPOSAL: The Color Classification step applies a classifier to pixel color rather than relying on mathematical color-space models or global image information.The paper uses Naïve Bayes and KNN as possible classification algorithms, with the detailed implementation using Naïve Bayes and KNN.
  • IV. OUR PROPOSAL: The Texture Classification step extracts local features from regular, similarly patterned regions detected automatically with superpixels.Global image features could make small fire regions vanish in the image context.
  • IV. OUR PROPOSAL: Region Merge retains pixels classified as fire by both color and texture classification in the final output image.The method hypothesizes that simultaneous classification by both approaches gives a higher chance that a pixel is actual fire.
  • IV. OUR PROPOSAL: BoWFire is modular, allowing different feature-extraction algorithms and classifiers, while its parameter count depends on the selected intermediate algorithms.

A. Configuration

The experiments use separate fire-detection and supervised training datasets containing both fire and fire-like non-fire imagery. The selected implementation combines Naïve Bayes, modified SLIC, YCbCr color, and uniform-pattern LBP with three parameters.

  • A. Configuration: The evaluation dataset contains 226 images: 119 with fire and 107 without fire, covering varied emergency situations and fire-like regions.Fire examples include buildings, industrial incidents, car accidents, and riots; non-fire examples include sunsets and red or yellow objects.
  • A. Configuration: The training dataset contains 240 images at 50×50-pixel resolution, with 80 fire images and 160 non-fire images.Non-fire training images also contain red or yellow objects.
  • A. Configuration: BoWFire uses Naïve Bayes with automatic discretization for pixel-color classification, modified SLIC in YCbCr space for superpixels, and uniform-pattern LBP for texture.
  • A. Configuration: The configured method needs only 3 parameters: Ksp, m, and K.The experiments use K = 11 and m = 40; more compact superpixels were observed to produce more regular regions and better texture representation.

B. Description of the experiments

The experiments examine superpixel count, color-only classification, and the combined BoWFire pipeline. They emphasize recovering fire pixels during color classification and reducing false positives when texture is added.

  • B. Description of the experiments: The parameter experiment varies the number of superpixels Ksp across 50, 100, 150, 200, 250, and 300.Its purpose is to evaluate how Ksp affects BoWFire performance.
  • B. Description of the experiments: Color Classification is evaluated primarily with Recall because the color step should recover as many fire pixels as possible.The study compares BoWFire’s color-based method with Celik, Chen, Rossi, and Rudz; false-positive rate is handled in the texture step.
  • B. Description of the experiments: BoWFire Evaluation measures the effect of combining color and texture, with reducing false positives without affecting overall performance as the central criterion.The study also evaluates state-of-the-art methods combined with Texture Classification and uses the best experimentally obtained Ksp.

VI. RESULTS AND DISCUSSION

BoWFire combines color and texture classification to reduce false positives in still-image fire detection. Across datasets and comparisons, texture generally improved precision and F1-score while slightly reducing recall, with Ksp = 150 selected as the best setting.

  • Superpixel sensitivity: On the fire dataset, varying Ksp produced precision around 0.8, recall around 0.65, and F1-score around 0.72.Performance increased slightly up to Ksp = 150 and then behaved similarly at larger values.
  • Superpixel sensitivity: Ksp = 150 provided the best overall balance, yielding strong F1-score while keeping FPR near its lowest value.The authors selected this setting for both the fire and complete datasets.
  • Comparative evaluation: Color Classification had the strongest overall color-only performance, while Rudz achieved the lowest FPR by discarding more true positives.Rudz had FPR below 9% across datasets, but Color Classification had better recall and F1-score.
  • Texture effects: Adding texture improved precision by up to 1.30× for Color Classification and Celik on fire images and 2.28× on the complete dataset.Rossi showed the largest precision gains: 4.43× on fire images and 5.65× on the complete dataset.
  • Texture effects: Adding texture reduced recall by up to 15% for most methods, while F1-score increased up to 69% on fire images and up to 65% on the complete dataset.The recall decrease resulted from discarding some true-positive pixels, whereas F1-score generally improved.
  • Texture effects: Texture classification reduced false-positive rates by up to 80% on non-fire and complete datasets without compromising overall performance.The result supports texture classification as a false-positive filtering step.
  • Comparative evaluation: BoWFire achieved similar recall to Celik without texture information but with a smaller false-positive rate.In ROC spaces for both fire and complete datasets, adding texture moved all methods toward lower FPR.
  • Qualitative evaluation: Visual examples show BoWFire discarding reflections and fire-like regions that other methods retained, including all false positives in a sunset image.BoWFire also avoided false positives in one high-false-positive emergency image.

VII. CONCLUSIONS

BoWFire detects fire in images with state-of-the-art-compatible performance while producing fewer false positives and using three intuitive tuning parameters. Future work will extend feature extraction and target smoke and explosions.

  • BoWFire achieved performance similar to state-of-the-art methods while producing fewer false positives.
  • The evaluation systematically compared BoWFire with four former works and demonstrated consistent improvements.
  • Color and texture information let BoWFire discard false positives using only image content, without temporal information.
  • BoWFire uses three parameters that are more intuitive to tune than those of previous methods.
  • Future work will investigate combining feature-extraction methods and extending BoWFire to smoke and explosions.
Loading 1506.03495v1…