In computer vision, accurately detecting objects of varying sizes is a fundamental requirement for a broad spectrum of applications, including autonomous vehicle navigation and medical imaging diagnostics. This research develops a size-adaptive Convolutional Neural Network (CNN) framework, coupled with the introduction of the Parameterized-Swish activation function, to refine object detection precision and operational efficiency. The size-adaptive framework is engineered to dynamically adjust the CNN’s configuration in response to the object size distribution encountered, leveraging statistical analysis and algorithmic decision-making to enhance detection performance across a spectrum of object sizes. The introduction of the Parameterized-Swish activation function, characterized by its dynamic parameters, allows for an adaptable response to diverse input patterns, significantly outperforming traditional activation functions by promoting faster model convergence and heightened detection accuracy. Implementing the proposed model leads to substantial performance gains: an 11.4% increase in mAP and 40.63% higher FPS for small objects; a 48.42% reduction in training time for medium objects with improved mAP; and for large objects, a 16.9% cut in training time and 76.04% quicker inference. Overall, the model achieves more than a 12% boost in detection efficiency and accuracy across various scenarios.