In　lossy　video　coding,　in-loop　filter　can　improve　the　image　reconstruction　by　reducing　compression　artifacts　and　distortions.　Recently,　several　CNN-based　in-loop　filtering　algorithms　are　proposed　to　improve　HEVC.　However,　something　deficient　is　still　limiting　their　capability　to　further　boost　the　coding　efficiency.　Specifically,　things　like　prediction　residuals　and　coding-unit　boundaries,　which　is　quite　relevant　to　the　compression　artifacts,　were　usually　neglected;　moreover,　the　loss　functions　adopted　in　these　methods　usually　haven＇t　been　regularized　properly.　To　improve　these　inadequacies,　we　propose　a　deep　learning　based　in-loop　filter　for　HEVC　to　improve　its　rate-distortion　performance.　Firstly,　prediction　residuals,　the　compression　of　which　is　actually　where　the　encoding　noise　directly　results　from,　are　skillfully　considered　to　help　filtering.　Then,　we　unify　a　block-based　measurement-reconstruction　process　and　a　neighborhood-based　filtering　process　into　one　end-To-end　all-convolutional　architecture　to　catch　the　relevance　between　compression　noise　and　the　pixel＇s　related　position　to　the　coding　block　boundaries.　Finally,　for　network　training,　total-variation　model　is　adopted　to　derive　the　loss　function　by　MAP　(maximum　a　posteriori)　estimation　for　better　regularization.　Experimentally,　our　proposed　in-loop　filtering　method　brings　average　6.5%　and　4.3%　BD-rate　reduction　under　AI　and　RA　configuration　respectively.　©　2019　IEEE.