This paper presents a novel borehole image compression algorithm for real-time (RT) logging while drilling (LWD). The compression scheme is designed to optimize the critical information required for RT decision making at low telemetry bandwidths. In the proposed algorithm we estimate the structure of the image (i.e. the amplitude and phase shift of the dip) and modify the encoding dictionary based on the features. The resulting dictionary resembles sinusoidal features, thus optimizing the reconstruction of bedding or other planar features in deviated wells. The dictionary is designed using a modified version of the 2D discrete wavelet transform (DWT). This approach has a low encoding complexity and supports the integration of directional information into the transform. Since feature estimation is a challenging step, we use a classifier to identify when directional information should be added to the transform or whether a conventional implementation is used. The algorithm has been implemented in both oil- and water-based mud LWD imager tools, where the low encoding complexity has facilitated the implementation in legacy tools with limited computation resources. We present field test results comparing the borehole images from RT and recorded mode (RM) data from one of the industry’s first RT LWD resistivity images obtained from a well drilled using oil-based mud.