Improved CNN Model Stability and Robustness with Video Frame Segmentation

This paper investigates the impact of image segmentation on improving the stability and robustness of convolutional neural network (CNN) models, particularly those that handle frames with tiny informative objects. Firstly, the authors introduced a new frame segmentation algorithm designed to preprocess video frames before they are subjected to classification. Secondly, it was proposed to use the average absolute difference between the accuracy of the training and the validation as a metric to measure the reliability and consistency of CNN models during training. This demonstrates the efficacy of the proposed techniques in augmenting image classification outcomes, particularly in scenarios where crucial objects constitute only a small segment of the frame, which allows one to solve similar problems in science. This research not only addresses specific challenges within the realm of sports footage analysis, but also offers broader implications for similar image classification tasks across various domains, thereby setting a foundation for future explorations in enhancing CNN model performance.