Development of reliable algorithms for the automatic diagnosis of broken rotor bars in induction motors (IM) has become the concern of many researchers during these previous decades. Though conventional steady-state current-based diagnosis approaches have behaved well for certain industrial applications, they may be not suitable in cases in which the machine does not operate under ideal stationary conditions (e.g. presence of load torque oscillations, supply unbalances, noises…). Due to this fact, alternative transient-based techniques based on the application of Time-frequency Decomposition (TFD) tools, have been introduced. They have shown satisfactory results, even in cases in which the conventional methodology does not work properly. Nonetheless, necessity of user expertness for the qualitative interpretation of the resulting time-frequency fault-related patterns as well as lack of automation in the diagnosis process makes often difficult their potential implementation in portable condition monitoring devices. A new algorithm for the automatic diagnostic of rotor bar failures is proposed in this paper. It takes as a basis the wavelet signals resulting from the Discrete Wavelet Transform (DWT) of the startup current, which contain basic fault-related features. These signals are further processed to generate 2-D images containing characteristic L-shaped patterns associated with the analyzed fault. Subsequent application of the scale transform allows obtaining scale-invariant feature matrices. Final correlation between these matrices enables to diagnose the condition of the machine. Test results prove the reliability of the algorithm and its generality to automatically diagnose the fault in machines with rather different sizes and load conditions.