Multiple-Input Multiple-Output (MIMO) systems play a crucial role in advancing 5 G and 6 G networks by improving spectral efficiency and data rates. In Frequency Division Duplex (FDD) systems, since the channel reciprocity does not hold, timely feedback of Channel State Information (CSI) from receiver to transmitter is essential to enable beamforming and precoding - which are fundamental techniques in MIMO systems. However, in MIMO configurations with large antenna arrays, this feedback mechanism introduces significant overhead to the system. Among different compression methods, deep learning based approaches outperform traditional compressive sensing techniques, but their large parameter count poses challenges for deployment on resource-limited devices, such as user equipment. Moreover, factors such as short channel coherence time and imperfect wireless channels demand more robust models. This paper introduces CLLWCsiNet (Convolutional Latent-space Low Weight CsiNet), a novel architecture that employs convolutional latent-space learning under imperfect channel conditions. CLLWCsiNet demonstrates superior predictive accuracy in noisy environments. We provide a comprehensive overview of previous models and introduce convolutional quantization techniques, which significantly reduce parameters and enhance computational efficiency. Our approach effectively compresses CSI while maintaining predictive accuracy, making it well-suited for deployment on resource-constrained user equipment. Simulations show that CLLWCsiNet achieves -11 dB Normalized Mean Square Error (NMSE) in low-SNR regimes (0-20 dB), achieving 31% lower NMSE than DNNet with only 70 K parameters.