OPTIMIZATION OF THE ANNOTATION PROCESS FOR BIOLOGICAL OBJECT IMAGES USING COMPUTER VISION METHODS

Abstract

This study presents an approach to the automated creation of an annotated dataset containing images of biological objects, particularly cells. The proposed methodology is based on a modified CRISP-DM framework, adapted to the specifics of computer vision tasks. A sequence of stages and steps has been developed to enable effective detection and localization of biological objects in microscopic images. The process involves preprocessing the images, including binarization, filtering, brightness and contrast adjustment, as well as correction of illumination artifacts. These operations help enhance the quality of the input images and improve the accuracy of subsequent detection steps. Detected objects are automatically localized based on morphological analysis, followed by clustering using the k-means algorithm. Grouping is based on features such as object size and mean color value, which allows for distinguishing between different types of cells or structures based on visual characteristics. Bounding boxes are automatically generated for the localized objects, and their coordinates are stored in a structured tabular format (.csv). The resulting dataset can be used to train or test deep learning models, particularly for tasks such as object localization, classification, or segmentation. The proposed approach was validated using images of blood smears containing various types of cells. All computations were carried out using the Python programming language and libraries such as Pandas, NumPy, OpenCV, and Matplotlib. The analysis of detection and classification accuracy demonstrated satisfactory results, confirming the feasibility of using the developed pipeline for automated generation of annotated biological image datasets.