Capsule Endoscopy has emerged as a non-invasive diagnostic tool for gastrointestinal diseases, yet efficient disease classification remains a challenge due to inherent complexities in image analysis. Also, the extensive time required for manual examination of capsule endoscopy led researchers and clinicians to seek time-efficient automated detection methods. This is where the profound advantages of Deep Learning (DL) become crucial. This research proposes a novel approach combining L-Softmax with Laplacian Smoothing Stochastic Gradient Descent (LSSGD) in ResNet architecture to enhance disease classification accuracy in capsule endoscopy images from Kvasir dataset. The L-Softmax function is integrated within a DL framework, facilitating better class separation and feature representation. Additionally, LSSGD is employed to mitigate overfitting and enhance model generalization. Experimental results show that our methodology is stable and easy to utilize in capsule endoscopy.