Improving performance of convolutional neural networks using adaptive morlet wavelet filters

Currently, the proliferation of Convolutional Neural Network (CNN) variables is a direct result of the exigency for meticulous image analysis and classification. In the realm of traditional CNN architecture, the sole means of discerning nonlinear patterns resided in the utilization of the pooling layer and the activation function. Therefore, CNNs must augment the count of convolutions and pooling layers. Additionally, every kernel employs constant coefficients, enabling it to approximate just one pattern within the input image. Given the escalating intricacy and growing diversity of patterns, it becomes imperative to increase the number of filters and consequently, augment the count of CNN variables. The focus of this study is to incorporate Morlet wavelet functions into the kernels, aiming to bolster their capability to perceive and extract intricate patterns. Given the employment of non-constant coefficients in the suggested kernels, the coefficients within the kernel adjust proportionally to the input. Consequently, this amplifies the spectrum of diverse patterns that the kernel discerns within the image. Moreover, the effectiveness of the functions and the methodology implemented in the proposed Kernel holds the potential to substantially curtail computational expenditures, stemming from the utilization of Morlet kernels with non-constant coefficients. The proposed method proposes a significant reduction in the count of trainable parameters, marking a decrease of 98% in comparison to the conventional structure and an 83% decrease when contrasted with the Volterra convolutional method.