Improved Mathematical Hybrid Deep Learning Methods for Blood Cell Image Recognition on Embedded Platforms

In this research work, improved mathematical hybrid algorithms for object detection and segmentation from microscopic images were developed. In the study, object detection algorithms such as Haar Cascade, Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) were combined with Convolutional Neuron Network (CNN) and Deep Neural Network (DNN) to create hybrid systems such as Haar Cascade+CNN and Haar+DNN. In the improved approach, Haar Cascade with dynamic threshold, multi-scale LBP and CNN algorithms combining features were used. These hybrid systems were tested on our own dataset, and their success rates were analyzed and compared. In the traditional method, CNN alone achieved 76.3% accuracy, while Haar Cascade+CNN achieved 79.6%. In the improved method, Haar Cascade+CNN reached 82-85% accuracy (average 83%). This means a 2.9-5.4% higher result than the traditional method. In terms of speed, the Haar Cascade+CNN method on Raspberry Pi5 was reduced from 0.130 seconds to ∼0.125 seconds. Tests were conducted on special computer systems such as Nvidia Jetson TX2 and Raspberry Pi5. The results show that the improved Haar Cascade+CNN method outperforms other hybrid methods in terms of accuracy and speed and provides high performance in microscopic cell image recognition.

Перевод пока недоступен