Weighted Average Ensembled Probability Pipeline of Pre-trained CNN Models for Hand Gesture Classification

The automatic detection of sign language from hand gesture images is fundamental for effective human-computer interaction, especially for individuals with hearing and speech disorders. Achieving accurate detection and classification of sign language gestures is paramount. However, accurately recognizing these gestures frequently presents challenges due to dull backgrounds, variations in gesture orientation, and fluctuations in lighting conditions. Previous approaches tried to overcome these challenges but still the model performance lacks recognition accuracy and visual quality. To address this need for accuracy, we proposed a Weighted Average Ensembled Probability Pipeline composed of EfficientNetB7, ResNet-50, and Xception models. Leveraging the Extensive Hand Gesture Recognition Database (HGRD), comprising 20,000 images, our individual models achieved notable accuracies: 96.498 % for EfficientNetB7, 93.2 % for ResNet-50, and 90.175 % for Xception. Through ensemble learning, we integrate these models to attain an impressive overall accuracy of 98.95 %. This research contributes to robust gesture recognition system and underscores the significance of ensemble techniques in enhancing performance for Gesture recognition systems.