This paper introduces an innovative system for image recognition using stacked intelligent metasurfaces (SIMs) and deep learning (DL) technologies. The proposed system leverages the diffractive neural network (DNN) architecture of SIMs to perform complex calculations at the speed of light, enhancing the efficiency, reliability, and security of information transmission. In this system, a SIM is positioned in front of the transmitting antenna, where the input layer of the SIM serves as a source encoder, and the multi-layer architecture forms a DNN for semantic encoding. The semantic encoder transforms signals into unique beams corresponding to the image class, which are then recognized by the receiver by probing the signal magnitude across the receiving array. An efficient mini-batch gradient descent algorithm is developed to train the transmission coefficients of the SIM's meta-atoms, achieving over 90% recognition accuracy in extensive simulations. The effectiveness of the SIM-based DNN for image recognition tasks is validated through various experiments, demonstrating its potential for practical applications in semantic communications.This paper introduces an innovative system for image recognition using stacked intelligent metasurfaces (SIMs) and deep learning (DL) technologies. The proposed system leverages the diffractive neural network (DNN) architecture of SIMs to perform complex calculations at the speed of light, enhancing the efficiency, reliability, and security of information transmission. In this system, a SIM is positioned in front of the transmitting antenna, where the input layer of the SIM serves as a source encoder, and the multi-layer architecture forms a DNN for semantic encoding. The semantic encoder transforms signals into unique beams corresponding to the image class, which are then recognized by the receiver by probing the signal magnitude across the receiving array. An efficient mini-batch gradient descent algorithm is developed to train the transmission coefficients of the SIM's meta-atoms, achieving over 90% recognition accuracy in extensive simulations. The effectiveness of the SIM-based DNN for image recognition tasks is validated through various experiments, demonstrating its potential for practical applications in semantic communications.