With　the　continuous　advancement　of　embedded　system　technology,　embedded　deployment　of　deep　learning　models　has　become　a　hot　research　issue.　Resource-constrained　devices　such　as　low-performance　microcontrollers　have　stringent　computational　and　memory　requirements,　so　how　to　design　efficient　and　lightweight　convolutional　neural　networks　(CNNs)to　adapt　to　the　limitations　of　these　devices　has　become　a　key　research　direction.　In　this　paper,　we　propose　a　lightweight　CNN　design　scheme　based　on　a　simplified　LeNet-5　[1]　architecture,　aiming　at　efficient　handwritten　digit　recognition　through　integer　weight　quantization　and　pure　C　programming.　The　network　is　trained　and　tested　on　the　MNIST　dataset,　and　the　optimized　model　is　successfully　deployed　on　a　low-performance　microcontroller　and　shows　significant　advantages　in　terms　of　computational　complexity　and　memory　footprint.　In　addition,　Ibuilt　and　implemented　this　lightweight　neural　network　using　pure　logic　gate　circuits.　Experimental　results　show　that　the　designed　lightweight　CNN　network　has　fewer　parameters　(1326),in　addition,　the　model　in　this　paper　has　a　lower　demand　in　terms　of　the　number　of　floating　point　operations　(920　FLOPs).　It　isable　to　achieve　fast　processing　and　significantly　reduce　the　hardware　burden　while　maintaining　reasonable　accuracy,　providing　an　effective　solution　for　deep　learning　applications　in　resource-constrained　environments.　©　2025,　John　Wiley　and　Sons　Inc.　All　rights　reserved.