DCT-Based White Blood Cell Image Enhancement for Recognition Using Deep Learning

White blood cell (WBC) recognition is still a challenging problem because of the high variability and complexity of blood cell images. Blood cell images can vary in quality, resolution, contrast, illumination, staining, and background. Blood cells can also vary in shape, size, color, texture, and distribution. Moreover, blood cells can overlap, cluster, or deform, making them challenging to segment and identify. This paper proposed an efficient automatic illumination compensation algorithm using singular value decomposition in the cosine domain (CSVDC) to enhance WBC images in the preprocessing step. Firstly, the WBC color image is split into three color channels and then mapped to the frequency domain using the discrete cosine transform (DCT) to get their DCT coefficient matrices. Next, the compensation coefficients are constructed based on the DC terms and DCT coefficient matrices’ most significant singular values. The DCT coefficient matrices are then linearly adjusted by multiplying with the compensation coefficients. Finally, three color channels are reconstructed using the inverted DCT to get the enhanced WBC image. Experimental results for the four most common PBC_dataset_normal_DIB, Raabin-WBC, BCCD, and Munich AML Morphology datasets using state-of-art deep learning models, including VGG16, GoogLeNet, and RestNet, illustrate the effectiveness of the CSVDC algorithm. In particular, on the PBC_dataset_normal_DIB dataset, when using the ResNet, the proposed enhanced WBC images have a higher average recognition rate compared to the original, ASVDF, ASVDW, and AHOSVD images by 3.82%, 2.69%, 11.37%, and 8.62%, respectively. Experimental results show that our method dramatically improves deep learning-based WBC recognition accuracy.