ILLUMINATION-AWARE DEEP LEARNING FOR WHITE BLOOD CELL CLASSIFICATION: A CBAM AND TEXTURE FEATURE FUSION APPROACH

Abstract

This study investigates the impact of illumination levels on classification performance in microscopic cell images and proposes a novel approach for the deep learning-based diagnosis of white blood cells. The proposed method integrates statistical texture features derived from grayscale images based on GLCM1 and LBP2 with the Convolutional Block Attention Module (CBAM) in a hybrid architecture built upon the VGG16 model. In this architecture, both visual (RGB) and structural (texture) information are processed simultaneously, aiming to obtain deeper and more meaningful representations for classification. Within this scope, the effect of varying illumination levels on microscopic images has been evaluated in detail for the first time, and the model's performance was tested under multiple lighting conditions. As part of the study, nine different illumination conditions were created by adjusting light levels from-80 to +80, and the model's classification performance under each condition was comprehensively evaluated. According to the findings, only 11