This paper presents a Discrete Cosine Transform (DCT) based Autoencoder (AE) and a Multilayer Perceptron (MLP) framework for Primary User (PU) spectrum data generation/ augmentation followed by sensing decision to develop cognitive radio networks (CRN). The proposed approach employs tailored DCT domain transformation to generate realistic and diverse spectrum snapshots, thereby enhancing the robustness of learning-based spectrum sensing (SS) under low Signal to Noise Ratio (SNR) and generalize Gaussian noise conditions. An AE is utilized to denoise and compress the spectral features, while a downstream MLP performs binary classification to determine the presence or absence of PU activity. We benchmark the proposed method against two recent works on Time-Frequency Cross-Fusion Network (TFCFN) and Short Term Fourier Transform-Convolutional Neural Network (STFT-CNN), which incorporates cross-attention for enhanced spectral feature fusion. Evaluations using probability of detection (Pd), probability of false alarm(Pf ), accuracy, and inference latency indicate that the proposed DCT-AE-MLP framework provides strong robustness at very low SNRs while maintaining competitive computational complexity. Specifically, DCT-AE-MLP achieves Pd ≈ 0.98 at Pf = 0.05, maintains Pd ≈ 0.90 at Pf = 0.3, and sustains Pd ≈ 0.72 at Pf = 0.7; across SNR levels, it achieves Pd ≈ 0.99 at 0 dB, 0.97 at -10 dB, 0.93 at -15 dB, and 0.84 at -20 dB.