A Novel Self Learning Activation Function-LogCosh Minimization Based Object Detection Model

Predicting accurate regions of interest in medical images, such as MRI, is paramount for effective diagnosis. This paper introduces an innovative object detection model adept at excelling in both normal and rotated image scenarios. By employing logcosh minimization, the model demonstrates enhanced accuracy in predicting bounding boxes. Moreover, it integrates novel activation function based on Gegenbauer polynomials, specifically tailored for small object detection, particularly in medical contexts. These activation functions possess the ability to learn and adapt their parameters during the training process, thereby better fitting the data and improving model performance. Through comprehensive experimentation, the proposed model showcases superior performance compared to the established object detection framework YOLOv5.