研究目的
To analyze the influence of different normalization methods on the discriminating ability of features estimated from the Gray Level Co-occurrence Matrix (GLCM) for texture classification.
研究成果
Normalization improves classification accuracy using GLCM-based features. The ±3σ normalization method is most effective for mitigating various image distortions, especially when the image histogram resembles a Gaussian distribution. Reduction in quantization levels can reduce classification error for additive noise but has limited effect for Rician noise.
研究不足
The study is limited to specific types of distortions (nonuniformity, Gaussian noise, Rician noise) and datasets (Brodatz textures and MRI phantoms). Further tests with combinations of distortions and other medical imaging modalities are needed.
1：Experimental Design and Method Selection:
The study involved texture analysis using GLCM features with various normalization methods (min-max, 1-99%, ±3σ) and quantization levels (3 to 8 bits per pixel). Linear discriminant analysis was used for classification.
2：Sample Selection and Data Sources:
Brodatz textures (12 images) and real magnetic resonance images (phantoms and brain tissues) were used. Brodatz images were corrupted with distortions: intensity nonuniformity, Gaussian noise (σ=25.6, 51.2), and Rician noise (s=0.1, 0.2).
3：6, 2), and Rician noise (s=1, 2).
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: Siemens Magnetom 1.5-Tesla scanner for MRI data acquisition; MaZda and QMaZda software for texture analysis.
4：5-Tesla scanner for MRI data acquisition; MaZda and QMaZda software for texture analysis.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Images were normalized and quantized. GLCM features were extracted for 4 directions and 5 offsets. Classification was performed using a binary linear classifier with feature normalization.
5：Data Analysis Methods:
Classification error was calculated based on confusion matrices. Results were plotted against quantization levels for different normalization schemes.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
