The growing need for fast and efficient processing by Machine Learning algorithms in the face of the exponential increase in data in various areas of activity has driven the search for innovative solutions. A promising approach is the use of light, to take advantage of increased parallelism, large bandwidths and low energy consumption inherent to optical methods. In this context, we present here a hybrid optical-electronic classifier that employs speckles designed to generate on-demand intensity distributions, following normal or non-normal statistics as needed. Using the Extreme Learning Machine technique, the system is trained to recognize image properties encoded in speckle patterns. In numerical simulations as well as in experiments, we observed a significant variation in image classification accuracy when transitioning from normal to non-normal regime. The results show satisfactory performance of the system for handwritten digit recognition.