This study presents a novel method for constructing class-oriented correlation filters, utilizing a unique integration of representation learning with phase-only autoencoders. Our proposed model employs a specialized autoencoder architecture designed to handle phase data effectively. The network is composed of dense layers and convolution layers, arranged to optimally condense phase information into a compact, low-dimensional representation, which is then used to reconstruct the input while preserving essential phase relationships. This processed information from the latent space is exploited to develop a correlation filter that is specifically tuned for class recognition, enhancing discriminant phase-encoded features. To further refine the classification threshold, kernel density estimation was employed, allowing for an empirical determination of decision boundaries based on the density functions derived from high and low peak to sidelobe ratio values. We rigorously evaluate our approach across benchmark datasets , YaleB and PIE, demonstrating its superiority in classification accuracy over state-of-the-art frequency domain and feature extraction methods. The effectiveness of our approach is affirmed through extensive testing on face recognition under varying lighting conditions.