Comparison of the Effectiveness of Using Various Approaches in Detecting Objects on Low-Quality Images

Machine methods of image analysis are gaining popularity in various fields of life. However, the question remains as to how effective such algorithms are on low-quality data, such as those that can be used in the field of telemedicine. The work provides a comparative analysis of various approaches to object detection in MRI brain images taken from a computer screen. For the recognition of brain contours in the image, a classical morphometric approach (OpenCV library), the Viola-Jones algorithm, and two deep learning algorithms, YOLOv8 and EfficientDet, were used. The comparison of these methods was conducted in terms of the quality of object detection in the image. To assess the quality, we used the IoU metric, as well as measured the amount of memory used and the speed of algorithm execution. As a result of the comparison, we found that the YOLOv8 model demonstrated the best performance in terms of object detection quality. However, its performance was unstable in cases of low-quality images with high levels of noise. Among the considered approaches, YOLOv8 is also the most memory-intensive. The YOLOv8 network architecture can be considered the best candidate for further practical application in terms of average performance and resistance to noise.