**Abstract**

The number of applications requiring effective object detection and classification are constantly increasing. In this regard, one of the most utilized solutions are Convolutional Neural Networks (CNN), which are also constantly pushed to improve their accuracy. When improving CNN, the key layers are the ones related to feature extraction and size reduction. In recent years, one way to improve the layers behavior is to combine them with different techniques, like the Multi-Resolution Analysis (MRA). The aim of the study is to use the Wavelet Transform Coefficients randomly incorporating the Lifting Scheme, which is a second-generation of MRA in a formal model as a pooling layer inside the CNN. This can emulate the Max-pooling dynamicfilter effect that prevents the CNN from overfitting but independently from the signal's shape, achieving a measurable increment of its accuracy using a benchmark dataset. To validate this study, seven pooling methods, including the proposed model, are tested using three different wavelet functions and one benchmark dataset. The results are compared with five of the most used pooling methods, which are also considered as the state-of-the-art.

**Keywords:** convolutional neural network, feature extraction, lifting scheme, multi-resolution analysis, random pooling, wavelet coefficients
