Abstract
In this research, we introduce a unique CNN model that stands out by using specialized layers to pick out the most important data for recognizing features. We focus on reducing unnecessary data and increasing data efficiency, which lets our model understand the data's layout better. This lets us capture a wide range of features in detail. We also address the common issue of SoftMax Loss causing overfitting by adjusting the adaptive margin in SoftMax. Our method not only makes feature types more varied but also groups similar items closer, which helps in distinguishing between different classes better. Plus, we've made the Adam optimizer better to ensure faster learning and to avoid getting stuck on local minima by incorporating an exponential weighted moving average and enhanced bias correction. Our algorithm’s performance is proven to be better than current standards, with very high accuracy in recognizing images across several datasets. The efficiency of the proposed framework is demonstrated on emotion detection tasks, outperforming existing benchmarks with an impressive 93.30% accuracy on the FER-2013 dataset, 91.94% on RAF-DB, and 100% on the CK+ dataset.