PaperlayerView exact revision on arXiv

Source-linked AI summary

Texture Image Classification Using DWT AlexNet Feature Fusion and Deep Neural Networks

Arun D. Kulkarni

arXiv:2608.28524v1cs.CVcs.AI

TL;DR

Texture classification needs both detailed multiscale spatial-frequency information and high-level discriminative representations, while handcrafted and deep features each provide only part of this coverage. The paper combines DWT and AlexNet features through feature-level fusion and classifies them with a DNN, with improved performance reported particularly for FMD, though gains vary by dataset.

  • Problem

    Texture classification requires multiscale spatial-frequency information and complex visual-pattern representation, which handcrafted features and CNN representations may capture only partially.

  • Method

    DWT spatial-frequency features and deep features from a modified AlexNet are extracted in parallel, concatenated into a unified representation, and classified using a DNN with a SoftMax layer.

  • Results

    Feature fusion improves performance particularly for the FMD dataset, while its effectiveness varies across datasets and can be comparable to DWT-DNN on Brodatz.

  • Takeaways & Limitations

    The framework provides a flexible approach for integrating handcrafted and learned features, with benefits depending on dataset diversity, complexity, and feature complementarity.

Abstract

from arXiv · show

Texture image classification plays a significant role in computer vision applications, including industrial inspection, medical image analysis, remote sensing, and object recognition. Handcrafted features can capture local texture characteristics but may have limited capability to represent complex visual patterns. In contrast, deep learning models automatically learn discriminative representations but may not fully exploit the multiscale spatial-frequency information inherent in texture images. This paper proposes a hybrid feature fusion framework, termed DWT_AlexNet_DNN, which combines Discrete Wavelet Transform (DWT) features with deep features extracted using AlexNet for texture image classification.

I. INTRODUCTION

Texture classification depends on representations that capture local, multiscale patterns while also modeling high-level visual characteristics. The paper addresses complementary limitations of handcrafted and CNN features through DWT–AlexNet feature fusion followed by DNN classification.

  • Texture classification supports applications including industrial inspection, medical image analysis, remote sensing, material recognition, and object recognition.
  • Handcrafted descriptors provide explicit local and frequency-domain texture representations but may have limited capability to represent complex, high-level visual patterns.
  • CNNs such as AlexNet learn hierarchical discriminative features, but may not explicitly preserve detailed multiresolution spatial-frequency information.
  • The proposed DWT–AlexNet–DNN framework extracts DWT spatial-frequency features and AlexNet hierarchical features in parallel, concatenates them, and classifies the fused vector with a DNN and SoftMax output.
  • The framework is evaluated on Brodatz, KTH-TIPS, and FMD using 70% training, 15% validation, and 15% testing subsets, with comparisons against individual feature representations.
  • Combining learned deep representations with frequency-domain information is presented as a straightforward alternative to specialized wavelet-CNN architectures.

III. PROPOSED APPROACH

The proposed approach evaluates DWT-only, AlexNet-only, and fused feature configurations for texture classification. In the fused configuration, independently extracted representations are concatenated and classified by a DNN.

  • III. PROPOSED APPROACH: The DWT-DNN model extracts DWT-based texture features and classifies them using a DNN.
  • III. PROPOSED APPROACH: The AlexNet-DNN model uses a modified AlexNet to extract a compact deep feature representation before DNN classification.
  • III. PROPOSED APPROACH: The DWT-AlexNet-DNN model performs feature-level fusion by concatenating DWT and modified-AlexNet features, then classifies the unified representation with a DNN.

A. DWT-DNN MODEL

The DWT-DNN model uses a three-level two-dimensional wavelet decomposition to derive statistical descriptors from localized spatial-frequency sub-bands. These descriptors form a 192-dimensional vector supplied to a fully connected DNN with SoftMax output.

  • A. DWT-DNN MODEL: A three-level two-dimensional DWT decomposes each image into approximation, horizontal, vertical, and diagonal detail sub-bands at each level.
  • A. DWT-DNN MODEL: Each decomposition level produces 64 features from four sub-bands and 16 statistical descriptors per sub-band.4 × 16 = 64 features per level.
  • A. DWT-DNN MODEL: 192 DWT features are supplied to a DNN consisting of fully connected layers followed by a SoftMax output layer.The total is 3 × 4 × 16 = 192 features.

B. ALEXNET-DNN MODEL

The AlexNet-DNN model uses modified AlexNet to extract a compact 256-dimensional deep feature vector, which is classified by a DNN with SoftMax output. The hybrid framework additionally concatenates DWT and AlexNet features before DNN classification.

  • AlexNet feature extraction: Modified AlexNet extracts a 256-dimensional feature vector before the final SoftMax classification layer.Convolutional layers learn progressively higher-level feature maps, while fully connected layers provide the compact representation.
  • AlexNet-DNN baseline: The AlexNet-DNN baseline uses learned deep features independently of DWT features to assess their discriminative capability.A SoftMax layer assigns each image to one of the ten texture classes during training.
  • Feature fusion: DWT captures multiresolution frequency characteristics, whereas AlexNet learns higher-level spatial and visual patterns from the same image.These complementary representations can be concatenated at the feature level.
  • Feature fusion: The fused representation contains 448 dimensions, combining 192 DWT features with 256 AlexNet features.The concatenated vector is supplied to fully connected layers followed by a SoftMax classifier.
  • Training and classification: The predicted class is the class with the largest posterior probability from the SoftMax output.The classifier input is the fused feature vector, and C denotes the number of texture classes.
  • Training and classification: Categorical cross-entropy training encourages high probability for the correct texture class and lower probability for incorrect classes.The average loss is computed across samples and classes.

PROTOCOL

All three models use the same 70%–15%–15% training, validation, and testing partition. Validation supports monitoring and model selection, while the independent test set is reserved for final evaluation using four classification metrics.

  • Data partitioning: The dataset is partitioned into 70% training, 15% validation, and 15% testing subsets for fair model comparison.The training set optimizes parameters, and the validation set monitors training and supports model selection.
  • Data partitioning: The independent test set is used only after training for final performance evaluation.This protocol separates model development from final assessment.
  • Evaluation metrics: Performance is evaluated using accuracy, precision, recall, and F1-score.F1-score is the harmonic mean of precision and recall.

E. OVERALL PROPOSED FRAMEWORK

The proposed framework processes each texture image through parallel DWT and modified AlexNet branches, concatenates their outputs, and classifies the resulting representation with a DNN–SoftMax classifier. The three configurations enable empirical comparison of the complementary information contributed by the two feature-extraction techniques.

  • Overall framework: The DWT branch produces 192 features and the modified AlexNet branch produces 256 features for each input texture image.The branches operate in parallel on the input image.
  • Overall framework: The two feature vectors are concatenated into a 448-dimensional representation before DNN–SoftMax classification.The overall mapping runs from the input image through DWT and AlexNet extraction, fusion, and classification.
  • Comparative evaluation: Comparing DWT-DNN, AlexNet-DNN, and DWT-AlexNet-DNN measures the complementary information contributed by the two feature-extraction techniques.Improved fused-model performance would indicate complementary information for texture discrimination.

IV. IMPLEMENTATION AND RESULTS

The evaluation compares DWT-DNN, AlexNet-DNN, and fused DWT-AlexNet-DNN representations across benchmark texture datasets, beginning with Brodatz.

  • Implementation: Three MATLAB classification models compare DWT-only, AlexNet-only, and fused feature representations.The fused model concatenates 192 DWT features with 256 AlexNet features into a 448-dimensional vector.
  • Brodatz Dataset: Brodatz contains 54 grayscale texture images spanning natural and man-made surface patterns.The dataset includes fabrics, wood, sand, leather, stone, and other materials.
  • Brodatz Dataset: DWT features from Brodatz images are clustered into granular, coarse, and patterned classes using K-means.The three clusters define the classification classes used in the experiments.
  • Brodatz Results: The Brodatz evaluation includes sample images, training progress, confusion-matrix analysis, classification results, and comparison of the DWT-based models.The text reports comparable performance for DWT-DNN and DWT-AlexNet-DNN across evaluated metrics.

B. KTH-TIPS DATASET

KTH-TIPS evaluates texture classification under changing illumination, pose, and scale using real-world material images distributed across ten classes.

  • Dataset Characteristics: KTH-TIPS contains 810 images across 10 texture classes, with 81 images per class.Images vary in illumination, scale, and viewing angle, creating within-class variation and between-class local-pattern similarity.
  • Dataset Characteristics: The dataset is used to assess robustness and generalization of texture descriptors and classification models under imaging-condition variation.Its real-world material images provide a challenging benchmark for handcrafted and deep-learning methods.
  • Evaluation: The KTH-TIPS section presents sample images, model training progress, confusion-matrix analysis, and classification results.These materials correspond to Figures 8–11 and the KTH-TIPS results table.

C. FLICKR MATERIAL DATASET (FMD)

FMD benchmarks material and texture recognition using diverse real-world Flickr images captured under varying illumination, viewpoint, scale, and background conditions.

  • Dataset Characteristics: FMD covers material categories including fabric, glass, leather, metal, paper, plastic, stone, and wood.It is a benchmark for material and texture recognition in computer vision.
  • Evaluation: The FMD evaluation includes sample images, training progress, a confusion matrix, classification results, and a dedicated results table.These components are represented by Figures 12–15 and Table V.
  • Dataset Characteristics: FMD exhibits substantial intra-class variability and interclass similarity because images vary in illumination, viewpoint, scale, and background.These properties make it suitable for evaluating robustness and generalization.

V. CONCLUSIONS AND FUTURE WORK

The paper concludes that DWT and AlexNet features provide complementary texture information, while their fusion is more useful for diverse datasets than homogeneous ones. Future work targets broader architectures and datasets, compact representations, and practical scalability.

  • Conclusions: The DWT-AlexNet-DNN model combines DWT descriptors, a 256-dimensional AlexNet vector, and a DNN classifier using a 448-dimensional fused representation.DWT-DNN and AlexNet-DNN provide comparison models for quantitatively assessing feature fusion.
  • Results: KTH-TIPS accuracy rises from 99.17% with AlexNet-DNN to 100% with DWT-AlexNet-DNN.For Brodatz, DWT-DNN and the fused model achieve comparable performance across evaluated metrics.
  • Conclusions: DWT descriptors capture multiresolution frequency-domain information, while AlexNet supplies higher-level discriminative representations.The paper characterizes these representations as complementary across texture images.
  • Conclusions: Feature fusion improves performance particularly for FMD, but its effectiveness varies with image diversity, complexity, and representational complementarity.Homogeneous datasets may already contain sufficient discriminative information in a single representation.
  • Future Work: Future work will test other CNN architectures and larger, more diverse datasets while investigating feature selection and dimensionality reduction.The goal is to make fused representations more compact while maintaining or improving classification performance.
  • Future Work: Future studies will examine computational complexity, scalability, and generalization to unseen datasets and imaging conditions.These evaluations are intended to clarify suitability for real-world texture classification applications.
Loading 2608.28524v1…