Enhancing recognition accuracy and efficiency through intelligent frame selection in uncontrolled conditions

This Paper explores the original approach to building pattern recognition systems with built-in modules for assessing the quality of input images and feedback at each stage of processing. This approach provides the ability to control the confidence of the result and failure in further processing in case of low quality of the input image at the considered stage, which allows to increase the accuracy, stability and speed of the recognition systems in uncontrolled shooting conditions. The necessary definitions have been introduced and constructs a model for the recognition system, paving the way for software implementation within both existing and newly designed systems. The significance of this approach is demonstrated through an example: selecting the best frames for recognition in a video stream. This selection is based on a frame priority function that considers image quality and shooting time.

The paper presents experimental results for a system that recognizes instruments in a video stream captured with automatic camera focusing, which often leads to blurred frames. These results show a significant increase in recognition accuracy, even with limitations on the video stream duration, when frames are selected based on their priority.