A vision-based tactile sensor is designed to satisfy the need of texture classification when the robotic hand grasping an object. The sensor includes three parts: a slab of elastomer, lighting devices and a camera. After obtaining clear images from this sensor, invariant LBP algorithm is used to determine which type the texture belongs to. In order to accelerate the computation speed and widen the use range, the algorithm is transplanted to circuits and verified in FPGA. Tests showed that the sensor can recognize 12 types of texture, including fingerprints, leather, fabric and so on. It can also get a real time result with the speed of 24fps.