Hyperparameter optimization (HPO) is crucial for enhancing Convolutional Neural Networks (CNNs) in tasks like image classification. Traditional methods like grid and random search are computationally intensive, while standalone approaches like Particle Swarm Optimization (PSO) and Bayesian Optimization (BO) face challenges such as premature convergence or scalability in high-dimensional spaces. To address these issues, a novel hybrid algorithm integrating Stochastic Particle Swarm Optimization (SPSO) with Bayesian Inference is proposed. SPSO introduces stochastic noise, enabling particles to escape local minima and maintain diversity, while Bayesian Inference prioritizes promising regions in the hyperparameter space for efficient exploration and exploitation. Evaluated on the CIFAR-10 dataset using SimpleAlexNet, the hybrid SPSO + BO algorithm significantly improved classification accuracy with fewer evaluations compared to standalone PSO and BO, demonstrating robustness, scalability, and reduced computational overhead. This work highlights the potential of combining stochastic and probabilistic optimization techniques for effective deep learning hyperparameter tuning.