Impact of Handcrafted Features and Raw Image Data on Neural Network Performance: Insights from the Brodatz Texture Database

Key computer vision tasks include image retrieval and classification, which are used extensively in fields including multimedia, security, and healthcare. This study contrasts conventional deep learning methods combined with feature extraction techniques.The study extracts feature by combining Completed Local Binary Patterns (CSLBP) and Grey Level Co-occurrence Matrix (GLCM), which are then put into a Dense Neural Network (DNN). CSLBP captures local texture patterns with reduced complexity, while GLCM analyzes the spatial relationships and co-occurrence of pixel intensities, enabling a robust representation of image features. Convolutional neural networks, or CNNs, are used in parallel for direct image classification. In comparison to the CNN model, which achieved data accuracy of 69.46%, the DNN with handcrafted features produced a test accuracy of 75.71% along with improved recall, precision, and F1-score values, according to the evaluation of both models. Plots of accuracy and loss throughout training and validation are also included in the study to show how the two models converge. These findings indicate that a DNN using handcrafted features performs better than a CNN and highlights the efficacy of conventional feature extraction techniques in image classification applications.