Mode-division multiplexing technology leveraging diverse spatial modes, has advanced to sustain capacity expansion in fiber-optic communications. The intelligent recognition of spatial modes using ultra-compact devices and low-complexity designs is crucial for mode visualization and system miniaturization. In this work, we theoretically design and experimentally demonstrate a neural network-optimized metasurface capable of dual-mode pattern recognition through dual-channel image display. Our framework offers three key advantages: device compatibility, design flexibility, and function scalability by integrating neural networks and metasurfaces into mode-division multiplexing platforms. Our framework enhances research and applications in intelligent metasurface-driven pattern recognition and object classification, information encoding and decoding.