Visual hand gesture detection is an interesting research field with an extensive range of applications like touch-less user interfaces, sign language translation, virtual and augmented reality interactions, remote control, etc. Hand gesture detection systems use various discriminative spatiotemporal attention features by calculating the dependencies between the finger and palm joints. However, these methods may face difficulties in achieving high performance and generalizability due to their inefficient features. Experimental results show that the approach presented in this paper achieves a state-of-the-art performance on a challenging dataset that is hand gesture recognition dataset, when compared to other published approaches. A novel hand gesture recognition model, Gestimage, is presented and rigorously evaluated against two state-of-the-art baselines that are EfficientNet and a Convolutional Vision Transformer (ConViT) across three benchmark datasets. These datasets vary in the number of gesture classes, image dimensions, and color channels, providing a comprehensive test of model robustness and generalization. On the four-class “Hand Gestures for Human–Robot Interaction” dataset, Gestimage achieves a test accuracy of 88.09%, surpassing EfficientNet (50.01%) and ConViT (64%). On the eight-class “Gestures Hand” dataset, Gestimage attains a test accuracy of 99.62% with a perfect F1-score of 1.00, outperforming EfficientNet (94.62%) and ConViT (98.25%). Notably, on the twenty-class “Hand Gesture Recognition” dataset, Gestimage reaches 100% validation and test accuracy, along with flawless precision, recall, and F1-score, significantly exceeding the performance of EfficientNet (91.78%) and ConViT (91.73%). The consistently high performance across ten evaluation metrics supports Gestimage as a highly effective framework for reliable and efficient hand gesture recognition in human–computer and human–robot interaction applications.