A Hybrid Multi-Level Feature Representation Framework for Improved Image Classification

Abstract

Image classification often requires representations that can capture both local texture details and higher-level visual information. In this paper, we propose a hybrid multi-level feature representation framework for better image classification by combining the handcrafted and deep features through a sequential processing pipeline. The local texture information is extracted using Local Binary Pattern (LBP) and the high-level features are extracted from pre- trained Xception model. Random Forest is then used to select informative features from both branches. The selected features are converted into compact mid-level representations using Bag of Visual Words (BoVW) with different vocabulary sizes. The resulting representations are fused and further refined using Apriori-based feature association mining before classification with a 1D ResNet-based classifier. The proposed framework is evaluated on four concrete-image datasets: CODEBRIM, DIMEC-Crack, BCD, and CCIC. The results show consistently high classification performance across the evaluated feature-vector sizes. The best accuracy reaches 99.98% on CODEBRIM, 99.98% on DIMEC-Crack, 99.98% on BCD, and 100% on CCIC. The results also show that increasing the feature-vector size does not always improve performance, indicating the importance of selecting a compact and informative representation.