Falls pose risks to the health and safety of the elderly, especially those living alone. Wearable and non-wearable fall detection systems have been developed in existing research, but these approaches are often expensive and intrusive. Consequently, researchers are showing increased interest in fall detection systems that utilize Wi-Fi signals, which are cost-effective and non-intrusive. Existing Wi-Fi signal-based fall detection systems typically employ deep learning techniques. However, systems trained on large datasets often achieve lower accuracy due to overfitting or insufficient data diversity. A new fall detection system using Wi-Fi Channel State Information (CSI) signals and deep learning has been developed to address this. The methodology involves the collection of Wi-Fi CSI signals using an ESP-32 microcontroller. Then, these signals are transformed into two-dimensional (2D) images using time series to image transformation techniques, the short-time Fourier transform (STFT), recurrence plots (RP), Markov transition field (MTF), and Gramian angular field (GAF). The resulting 2D images are fed into the proposed adaptive residual dense network with convolutional block attention module (ARD-CBAM) designed to extract features and improve recognition accuracy. The network's performance is further enhanced by optimizing its parameters using the Enhanced Namib Beetle Optimization Algorithm (ENBOA). Results show that the proposed ENBOA-ARD-CBAM scheme's F1 score is 10% better than the Resnet model and 5% better than the VGG-16 model. The region of convergence (ROC) of the proposed ENBOA-ARD-CBAM-based approach exceeds 0.8, outperforming other techniques with lower ROC values. Importantly, the developed ENBOA-ARD-CBAM provides a better balance between specificity and sensitivity, ensuring its reliability in fall detection.