In order to reduce the labor intensity and improve the feeding efficiency of thin and light packages, the work aims to design an intelligent feeding system that can replace manual labor. The Mask R-CNN instance segmentation model was trained based on a big dataset from production site, achieving accurate recognition of thin and light packages. The grabbing area of stacked thin and light packages was obtained by the technique of maximum inscribed circle in the package masks. Through the motion path planning of the robotic arm, smooth grabbing of thin and light packages was achieved. Double sided scanning of thin and light packages was achieved based on code readers and mirror reflection technology. The experimental results showed that the intelligent feeding system achieved accurate recognition and reliable grabbing and transporting of stacked thin and light packages, with a recognition accuracy of 90.02% and an average grabbing success rate of 96.82%. The intelligent feeding system constructed based on these methods can replace manual labor to efficiently complete the feeding of thin and light packages.