AGGREGATION OF PERCEPTRONIC PATTERN RECOGNITION RESULTS YOLO V8N PLATFORMS BASED ON THE MEDIAN METHOD

Abstract

Traditionally, the critical parameters of information systems, in particular robotic systems, include real-time signal and data processing. The paper considers modern platforms and methods for identifying objects from their two-dimensional images based on perceptron structures, in particular, the implementation of YOLOv8 is associated with a number of problematic issues, such as the need to optimize the model for limited resources and provide a high-quality training data set.

The main goal of the research is to improve existing approaches that can increase the accuracy of object detection with limited computing resources. The main attention is paid to technologies for pre-processing the output data of perceptron structures (which operate in parallel) with the subsequent use of smoothing to minimize the impact of noise.

As a result, a multi-model approach is proposed with parallel processing of one image by several (in this case three) perceptron structures of the YOLO platform. Aggregation of results is carried out by the method of selecting the median value of the identification result, which allows to reduce the impact of random errors and increase the stability of the system.

In addition, the advantages of the proposed approach are described, in particular: a typical approach, simplicity of implementation, resistance to errors, preservation of important features of object images. At the same time, its problematic aspects and limitations are noted, in particular, ignoring potentially important, but rejected predictions.

The study also considers the implementation of relevant algorithmic solutions, the features of their functioning, presented in block diagrams demonstrating the processes of processing by confidence level and selection of the most relevant results. Thus, the proposed approach to aggregating the results of several neural networks allows to improve the accuracy and stability of object identification in two-dimensional images, ensuring further effective integration. However, it should be understood that achieving the aforementioned advantages requires the involvement of several perceptron structures in parallel operation, which significantly increases the computational requirements for the implementation of such solutions.