Reinterpreting wavelet multi-resolution analysis as CNN methods to endow them with the capacity for high-level semantic feature extraction has emerged as a research topic in deep sparse representations. We explore the fundamental operations of wavelet transform and convolution from a projective standpoint, then design a highly interpretable, exceptionally lightweight, and fully learnable deep neural network architecture known as Depthwise Separable Axial Asymmetric Wavelet Convolutional Neural Networks (DSAWCN). This network acquires wavelets filters customized for specific image tasks, thus providing adaptive and efficient multi-scale feature extraction strategies. We examined the behavior and the impact of the parameters in the proposed method using three general texture image datasets and four bark texture image datasets. The findings indicate that this exceptionally lightweight network surpasses current wavelet convolutional networks in terms of classification accuracy and achieves performance comparable to some eminent CNN models.