Padding　is　used　to　maintain　the　size　of　the　feature　map　and　reduce　information　bias　against　boundaries.　The　extra　information　added　by　these　schemes　at　the　boundary　implicitly　affects　the　feature　extraction.　Zero-padding　introduces　weakly　correlated　information　resulting　in　weak　boundary　information,　but　provides　location　encoded　information.　Various　padding　produces　weight　asymmetries　due　to　the　uneven　application　in　subsampling.　With　the　lack　of　a　universally　superior　method,　it　is　necessary　to　manually　determine　whether　the　padding　method　is　suitable　for　a　given　task.　In　this　paper,　we　propose　a　self-learning　padding　mechanism,　S-Pad,　which　extends　the　value　by　learning　the　boundary　information　of　the　image　and　provides　the　network　with　a　set　of　1x1\documentclass[12pt]{minimal}　\usepackage{amsmath}　\usepackage{wasysym}　\usepackage{amsfonts}　\usepackage{amssymb}　\usepackage{amsbsy}　\usepackage{mathrsfs}　\usepackage{upgreek}　\setlength{\oddsidemargin}{-69pt}　\begin{document}$$1\times　1$$\end{document}　filters.　Following　the　training　process,　tailored　boundary　rules　adapted　to　the　specific　task　can　be　obtained.　S-Pad,　in　turn,　effectively　mitigates　the　weakening　of　boundary　information　as　well　as　weight　asymmetries.　Our　study　is　dedicated　to　probing　the　effectiveness　of　S-Pad　in　the　domains　of　object　detection　and　classification.　Through　our　validation　process,　we　establish　the　enhanced　efficacy　of　S-Pad,　resulting　in　an　overall　performance　improvement　for　both　tasks.