Deep invariant texture features for water image classification

Minglong Xue, Palaiahnakote Shivakumara, Xuerong Wu, Tong Lü, Umapada Pal, Michael Blumenstein, Daniel LoprestiSN 2020 · DOI

Abstract

Accurate classification of water quality from standard photographic images is a challenging task due to variations in lighting, camera angles, and the visual similarity between clean and polluted water. This paper presents a hybrid classification framework that integrates handcrafted invariant texture features with deep learning representations. By combining Scale Invariant Gradient Orientations (SIGO) and Gabor Wavelet Binary Patterns (GWBP) with features extracted from a pre-trained VGG16 model, the proposed method achieves high accuracy even on small, complex datasets. Experimental results on a curated dataset of 1,000 images demonstrate a binary classification F-measure of 0.96, significantly outperforming traditional end-to-end deep learning models that are prone to overfitting in data-scarce environments.

1. Introduction

Monitoring water quality is essential for environmental protection, yet automated classification remains difficult for "normal" camera imagery compared to remote sensing or underwater photography. Pollutants such as algae, oil spills, and rubbish introduce subtle textural changes that are often obscured by reflections or varying environmental conditions. While deep learning models like ResNet have revolutionized image classification, they typically require large-scale labeled datasets. In niche domains like water pollution detection, data is often limited, leading to overfitting in complex deep architectures. This study addresses these challenges by proposing a hybrid feature fusion approach that leverages the strengths of both classical computer vision and modern deep learning.

2. Methodology

The proposed system utilizes a multi-stage feature extraction pipeline coupled with a Gradient Boosting Decision Tree (GBDT) classifier. The core innovation lies in the synergy between robust handcrafted descriptors and high-level deep features.

2.1 Handcrafted Feature Extraction

To capture robust textural information, the framework employs three primary handcrafted techniques:

  • Directional Coherence (DC): Images are transformed into the HSV color space to isolate perceptual details. Eigenvalue analysis on structure tensors is then used to highlight dominant gradient directions, enhancing directional consistency.
  • Scale Invariant Gradient Orientations (SIGO): To ensure robustness against varying camera distances, Gaussian first-order derivative filters are applied across multiple scales. The most stable orientations are selected through a quantization process, providing scale invariance that deep models often struggle to learn from limited samples.
  • Gabor Wavelet Binary Patterns (GWBP): This descriptor combines the multi-resolution analysis of Gabor wavelets with the local spatial structure captured by Local Binary Patterns (LBP), creating a comprehensive texture representation.

2.2 Deep Feature Fusion

In parallel with handcrafted features, a pre-trained VGG16 network is used as a fixed feature extractor. A 1,000-dimensional vector is extracted from the network's upper layers. These deep features are concatenated with the SIGO and GWBP descriptors to form a unified feature vector that captures both low-level geometric properties and high-level semantic content.

2.3 Classification via GBDT

The fused feature set is processed by a Gradient Boosting Decision Tree (GBDT). GBDT was selected for its efficiency and its ability to handle imbalanced or high-dimensional feature sets more effectively than standard neural network classifiers or SVMs when training data is limited. It balances disparate feature types through iterative optimization, making it ideal for the "messy" data found in environmental surveillance.

3. Experimental Results

3.1 Dataset

The authors established a new benchmark dataset comprising 1,000 images (500 clean, 500 polluted) sourced from various web engines and existing research. The polluted category includes diverse hazards such as industrial waste, oil, fungi, and algae. The data was split into 75% for training and 25% for testing.

3.2 Performance Evaluation

The hybrid model demonstrated superior performance across multiple metrics:

  • Binary Classification: The model achieved an F-measure of 0.96 in distinguishing clean from polluted water.
  • Multi-Class Classification: In a more challenging 10-class scenario, the model reached an average F-measure of 0.57, significantly outperforming the next best existing method (0.52).
  • Ablation Study: Results confirmed that while VGG16 features are powerful, the addition of SIGO and GWBP significantly boosts accuracy and provides better invariance to rotation and scaling.

4. Discussion and Conclusion

The integration of classical computer vision descriptors with deep learning features provides a robust solution for environmental monitoring tasks where data is scarce. The SIGO and GWBP features compensate for the deep model's potential lack of geometric invariance when trained on small samples.

However, limitations remain regarding sensitivity to Gaussian noise and motion blur, which are common in real-world drone or handheld photography. Future research will focus on enhancing the architecture's robustness to background clutter and improving multi-class discrimination for visually similar pollutants.

Summary updated

BibTeX
@inproceedings{xue2020deep,
  author = {Minglong Xue and Palaiahnakote Shivakumara and Xuerong Wu and Tong Lü and Umapada Pal and Michael Blumenstein and Daniel Lopresti},
  title = {{Deep invariant texture features for water image classification}},
  booktitle = {SN},
  year = {2020},
  doi = {10.1007/s42452-020-03882-w},
}
Figure