EVOLUTION OF IMAGE SEGMENTATION METHODS: CLASSICAL ALGORITHMS AND DEEP LEARNING

Abstract

Background. Text image binarization is an important preprocessing task in document analysis, optical character recognition, and automated information extraction. The task becomes challenging when images contain blur, noise, low contrast, or background degradation.

Materials and Methods. This study compares classical and lightweight neural methods for binarizing degraded text images. The evaluation was conducted on a paired text binarization dataset comprising degraded input images and their corresponding clean binary target masks. The evaluated methods included simple global thresholding, Otsu binarization, adaptive thresholding, K-means clustering, Fuzzy C-means clustering, watershed segmentation, region growing, region splitting and merging, Canny edge detection, Laplacian of Gaussian filtering, and a U-Net model. Tunable parameters were selected on the validation subset, while final results were reported on the independent test subset. The methods were assessed using Intersection over Union (IoU, Jaccard index), Dice similarity coefficient (DSC), Pixel Accuracy (PA), Precision, Recall, and execution time.

Results and Discussion. Adaptive thresholding achieved the best performance among classical methods: IoU = 0.785, DSC = 0.877, PA = 0.961, Precision = 0.851, and Recall = 0.923. Simple global thresholding was the fastest method, with a runtime of 0.0015 ms/image, but lower Precision. K-means, Fuzzy C-means, Otsu, and region growing demonstrated high Recall values above 0.97, but lower Precision, indicating that degraded background regions were often classified as foreground. Edge-based methods produced lower IoU and DSC because they mainly detected character contours. The Tiny U-Net model achieved the highest overall quality: IoU = 0.959, DSC = 0.979, PA = 0.994, Precision = 0.979, and Recall = 0.979.

Conclusion. The results show that different binarization methods are suitable for different scenarios, depending on the required accuracy, runtime constraints, and available resources. Classical methods remain useful for low-latency processing, while lightweight neural models provide higher accuracy when sufficient computational resources are available. Pixel Accuracy should be interpreted carefully because background pixels dominate document images.

Keywords: text binarization, image segmentation, thresholding, clustering, U-Net, evaluation metrics.