RNN–LSTM-assisted entropy-based source–channel coding for robust image transmission

The widespread growth of multimedia applications across wireless networks has increased the necessity for robust image and video communications. Wireless image transmission over noisy communication channels suffers from degradation in reconstruction quality due to bandwidth limitations and channel impairments. Existing methods often failed to effectively integrate entropy-driven compression with adaptive feature modelling, limiting their robustness and efficiency. To address this issue, this research develops an effective technique using Deep Learning with entropy-based source-channel coding for image transmission. Initially, an input image is processed using a Graph Cut-based Region of Interest (ROI) detection method to identify visually significant regions. The image is then partitioned into ROI and non-ROI regions and encoded independently. Here, the ROI region is encoded using LZW coding, whereas the non-ROI region undergoes wavelet transformation, quantization, and holoentropy-based arithmetic coding for enhanced compression. The encoded streams are combined into a unified final bitstream. Subsequently, a proposed Recurrent Neural Network (RNN)-Long Short-Term Memory (LSTM) model -based entropy encoder further optimizes the bitstream before transmission through an Additive White Gaussian Noise channel. At the receiver, a joint source–channel decoder reconstructs the transmitted data, followed by decompression and ROI-guided image reconstruction. The developed RNN-LSTM + entropy encoding achieved superior performance, yielding a high Peak Signal-to-Noise Ratio of 46.64 dB, minimum Mean Absolute Error of 29.3% minimum Mean Squared Error of 5.6%, maximum Structural Similarity Index Measure of 92.5%, higher Multi-Scale Structural Similarity Index Measure of 95.4%, and minimum Learned Perceptual Image Patch Similarity of 6.2%. These results confirm the effectiveness of the proposed framework for robust image transmission.