Recent advances in deep learning, especially the use of Convolutional Neural Networks (CNNs), show promise for automating the detection of suspicious activity using image analysis, common CNNs face obstacles such as high computational time and weak generalization challenges. To improve the CNN model, Optimisation strategies like the Spider Wasp Optimiser (SWO) have been explored, but issues with consistency and convergence remain. Therefore, this paper introduced an Enhanced Spider Wasp Optimiser (ESWO) that incorporates roulette wheel selection, designed to better fine-tune CNN hyperparameters for automated face detection and recognition. The spider wasp Optimisation algorithm was enhanced using roulette wheel selection method instead of random selection method in the standard SWO. Roulette wheel selection assigns selection probabilistic based on individual fitness, favouring solutions with higher fitness values and thereby guiding the search toward more promising regions in the solution space. The resulting enhanced Spider Wasp Optimisation (ESWO) which formed ESWO-CNN was then used to optimize CNN settings for feature extraction. The study gathered face images from LAUTECH students, faculty of computing and Informatics and applied a pre-processing workflow including resizing, cropping, grayscale transformation, and enhancing brightness and contrast training. The ESWO-CNN was implemented with the preprocessed facial images in MATLAB R2023a. The performance of the formulated model was measured using sensitivity, specificity, precision, accuracy, false positive rate, computation time, and compared against existing Spider Wasp Optimised-CNN (SWO-CNN) approach. The results revealed that the developed ESWO-CNN model exhibited superior performance, attaining the sensitivity of 99.18%, specificity of 98.92%, precision of 99.07%, accuracy of 99.07%, and F1-score of 99.18%, with False Positive Rate of 1.08% and minimal computational time of 60.09 seconds. The SWO-CNN model showed notable performance enhancement with a sensitivity of 98.12%, specificity of 97.53%, precision of 98.12%, accuracy of 97.87%, F1-score of 98.12%, and a reduced FPR of 2.47%, taking 71.64 seconds to execute. In conclusion, the ESWO-CNN approach constitutes a highly effective, scalable, and efficient method for the detection of suspicious activity using image analysis. Its robust parameter Optimisation supports security monitoring, addressing both traditional and modern security challenges.

Keywords: Convolutional Neural Networks, parameter Optimisation, Spider Wasp Optimiser, Intelligent Deep Learning Model.