隨著影像應用的普及，靈活且高效的影像壓縮技術需求日益增加。傳統編碼器如 JPEG、JPEG2000 和 BPG 雖具穩定性，但面對複雜影像內容與現代應用場景時已逐漸顯現其限制。近年來，學習式影像壓縮（Learned Image Compression, LIC）憑藉深度神經網路展現出優越的壓縮效率，但大多數方法仍須針對每個壓縮率訓練獨立模型，缺乏彈性。本研究提出一種具可變壓縮率能力的影像壓縮架構，結合 Transformer 與 CNN 模型，並引入 λ 調控機制以支援單一模型在不同壓縮率下自適應調整。我們設計了動態模組與選擇性副特徵通道，有效提升模型在各種碼率下的率失真表現。實驗在 Kodak、CLIC 與 Tecnick 等資料集上進行評估，結果顯示本方法在 PSNR、MS-SSIM 及 BD-rate 指標上均優於多項現有方法，展現出良好的靈活性與實用性，適合作為未來可變位元率影像壓縮的解決方案。;With the growing demand for image-related applications, there is an increasing need for flexible and efficient image compression techniques. Traditional codecs such as JPEG, JPEG2000, and BPG offer stable performance but have shown limitations when dealing with complex visual content and modern usage scenarios. Recently, learned image compression (LIC) has demonstrated superior compression efficiency through deep neural networks; however, most existing methods require training a separate model for each compression rate, lacking adaptability. In this work, we propose a variable-rate image compression framework that combines Transformer and CNN architectures, and incorporates a λ-conditioned mechanism to enable a single model to dynamically adjust its behavior across different bitrates. We design dynamic modules and a selective side-channel pathway to improve rate-distortion performance under various compression levels. Experiments conducted on standard datasets such as Kodak, CLIC, and Tecnick show that our method outperforms existing approaches in terms of PSNR, MS-SSIM, and BD-rate, demonstrating strong flexibility and practicality, making it a promising solution for future variable-rate image compression systems.
