Facial emotion recognition (FER) systems have gained popularity due to their applications in various fields, including healthcare, cognitive science, video conferencing, and driver safety. However, recognizing facial expressions automatically from images presents challenges due to inter-person variability and the difficulty in accurately distinguishing between different emotions. The research investigates the application of a modified Local Binary Pattern (LBP) in facial emotion detection, focusing on the multilayer multi-block local binary pattern (mL-mB LBP). The process involves converting facial expression images into \(3 \times 3\) identical blocks, which are then used to generate center pixels using the LBP descriptor block-wise. The final layer is partitioned into \(3 \times 3\) blocks to generate central pixels. The study introduces a customized convolutional neural network (CNN)-based framework for emotion recognition, which uses the mL-mB LBP image data to accurately predict emotions. The CNN model is trained using multi-level multi-block images. The experimental study findings show that the accuracy of the CNN model is significantly improved when trained using multi-block multi-level images with all layers fused.