Research on Image Classification Method Based on Haar Wavelet Pooling and Probabilistic Mixture Model

In high-dimensional image classification tasks, conventional convolutional neural networks (CNNs) suffer from high-frequency feature loss in pooling layers and limited robustness of Softmax classifiers when handling complex data distributions. In this paper, we propose a novel classification framework that combines multi-scale Haar wavelet pooling with a Gaussian Mixture Model Classifier (GMMC), which can simultaneously retain high-frequency information and optimize multimodal feature distributions. During model training, we integrate Selective Multiscale Wavelet Pooling (SMWP) into the Expectation-Maximization (EM) algorithm to enhance frequency-domain features and jointly improve classification accuracy. Our approach achieved classification accuracies of 97.17% on CIFAR-10 and 99.98% on SVHN, outperforming the MaxPooling + Softmax baseline by 7.32% and 4.93%, respectively. This research proposes a promising framework for fine-grained medical image classification, with potential applicability in low-light image enhancement and cross-modal retrieval tasks.