With the exponential growth of data traffic, optical fiber communication systems require advanced methods to optimize transmission efficiency while preserving signal integrity. This paper proposes a novel Convolutional AutoEncoder (CAE)-based semantic communication framework for optical fiber networks with an adjustable compression rate, balancing bandwidth efficiency and transmission quality. The system integrates two specialized models: CAE-N, which provides adaptive semantic compression with configurable compression ratios ranging from 2:1 to 98:1, and CAE-1, which focuses on high-precision denoising to mitigate optical channel impairments. The CAE-N model selectively reduces redundant information while retaining essential semantic content, significantly lowering the data load without sacrificing meaningful transmission. The Enhanced Gaussian Noise (EGN) model is employed to accurately simulate optical fiber effects, ensuring realistic performance evaluation. Experimental results demonstrate that CAE-1 achieves a 28.9% improvement in Structural Similarity Index (SSIM) over conventional transmission methods, ensuring high-fidelity signal reconstruction. Furthermore, CAE-N maintains competitive transmission quality even at high compression ratios, demonstrating its capability to significantly enhance bandwidth utilization while preserving semantic integrity. In addition to using MNIST for foundational training, this study also utilized other datasets for training and validation, including medical datasets, in both high-definition and standard-definition formats. The results demonstrate that our model possesses strong generalization capabilities. New datasets require very few training epochs to achieve performance that closely approaches, or even slightly exceeds, the results obtained on MNIST. This work bridges deep learning-based semantic communication with optical fiber transmission, offering a scalable, efficient solution for next generation optical networks that optimally balances compression efficiency and signal quality, suitable for various scenarios including telemedicine, industrial quality control, urban nighttime surveillance, archival digitization, and low-bandwidth collaborative platforms.