Differentiable　architecture　search　(DARTS)　is　an　important　branch　of　neural　architecture　search　(NAS),　aiming　at　searching　the　optimal　network　structure　efficiently.　However,　the　backbone　and　FPN　of　the　detection　model　tend　to　suffer　from　over-parameterization　effects　during　the　search　process,　resulting　in　DARTS　failing　to　efficiently　search　the　detection　network　structure.　In　this　paper,　we　propose　an　efficient　differentiable　architecture　search　(EDARTS-DET)　with　backbone　and　FPN　for　object　detection.　Firstly,　the　joint　search　space　of　backbone　and　FPN　networks　is　designed　to　facilitate　DARTS　to　search　the　detection　network　as　a　whole　at　one　shot.　Secondly,　an　architecture　operation　mask　weight　sharing　mechanism　is　proposed　to　effectively　reduce　the　search　memory　occupation　and　computational　cost.　Finally,　an　attention-based　partial　channel　selection　strategy　is　introduced　to　select　important　channels　to　be　fed　into　the　search　space.　The　experimental　results　show　that　the　proposed　EDARTS-DET　achieves　lower　search　consumption　time　and　memory　utilization　in　the　COCO　dataset　and　the　Haier　appliance　disassembly　dataset　compared　with　other　SOT　A　methods.　Meanwhile,　the　searched　detection　networks　are　also　improved　in　terms　of　detection　performance,　verifying　that　the　proposed　EDARTS-DET　method　achieves　a　well-balanced　performance　and　efficiency.