Converge intra-class and Diverge inter-class features for CNN-based Event Detection in football videos

Football event detection in videos is very challenging, but challenges on the Penalty and the Free-kick, which have common visual elements, are severe and critical. The existence of common elements between two events causes the extraction of common and ineffective features in recognizing these two events. As a result, the error of recognizing and separating these two events is more than other events. In this paper, we present a new method for filtering the input data to converge the intra-class features and diverge the inter-class features to increase the classification accuracy. For this purpose, using the IAUFD Dataset, we have evaluated images for the Penalty and the Free-kick classes with the criterion of structural similarity. Based on the results, inappropriate images have been ignored according to the average value and standard deviation of each class of data. This filtration leads to ignore of ineffective and common features in the learning process. The results of the proposed method indicate an improvement in the accuracy of distinguishing between two Penalty and Free-kick events using a deep neural network and filtered training images compared to the deep neural network using all training images.