WBFVD: A UNet-Based Deep Learning Model for True-to-Color Image Colorization of White-Background Fruits & Vegetables Dataset

Color is a vital attribute in fruit and vegetable images, especially for applications in e-commerce and agricultural systems. Accurate colorization of grayscale images—particularly those captured with white backgrounds—remains a challenging task due to limited texture cues and the ill-posed nature of color reconstruction. In this paper, we introduce a UNet-based deep learning model for true-to-color image colorization tailored specifically to the White-Background Fruits and Vegetables Dataset (WBFVD), which we curate and release as part of this work. The dataset includes 1965 grayscale–color image pairs across 20 fruit and vegetable categories. Our approach integrates specialized skip connections and a novel color rebalancing mechanism to enhance chromatic accuracy and structural preservation. The model achieves superior results over traditional machine learning, CNN, and GAN-based colorization methods, with a peak PSNR of 26.7 dB, SSIM of 0.825, and hue-based color compliance of 98%. Additionally, we conduct ablation studies and comparative evaluations to demonstrate the robustness of our method. This work sets a strong baseline for domain-specific image colorization and highlights future directions for real-world generalization and subjective visual assessment.