Kolmogorov–Arnold Networks (KANs) have recently emerged as promising alternatives to traditional neural architectures by offering greater parameter efficiency and functional adaptability, particularly for structured data tasks. Despite their potential, the effectiveness of these methods in computer vision remains relatively unexplored. In this study, we introduce a lightweight hybrid deep learning model that integrates the convolutional feature extraction strengths of Convolutional Neural Networks (CNNs) with the non-linear representational power of KANs for image classification on the CIFAR-10 dataset. Our CNN+KAN, called LKANet, architecture captures high-level spatial features via compact convolutional layers and enhances decision boundaries through learnable spline-based activations in the classifier head. Benchmarking against a parameter-matched state-of-the-art PyramidNet variant reveals that CNN+KAN(LKANet) achieves a superior test accuracy of 89.16%, delivers faster inference speeds, and maintains a significantly smaller memory footprint (7.30 MB). These results highlight the potential of spline-based learning mechanisms to develop lightweight, high-performance models suitable for real-time and edge-level computer vision applications.