In　log　trade,　precise　measuring　log　volume　is　directly　related　to　the　economic　benefits.　Usually,　the　length　of　the　same　batch　of　wood　is　relatively　fixed,　but　the　shape　and　size　of　log　end　are　different.　Therefore,　the　variability　of　the　size,　shape　and　density　of　log　end,　and　the　sophisticated　scenarios　lead　to　great　challenge　for　log　detection　from　bundled　log　ends　images.　In　addition,　the　large　number　of　logs　in　the　bundled　log　ends　image　results　in　less　pixels　occupied　by　a　single　log,　and　it　is　essential　to　detect　the　small　log　ends.　In　order　to　accurately　detect　each　log　from　bundled　log　ends　image,　a　new　approach　of　log　detection　based　on　K-median　clustering　and　improved　You　Only　Look　Once　(YOLO)v4-Tiny　network　has　been　developed.　The　specific　methods　are　as　follows:　(1)　In　consideration　of　the　characteristics　of　log　size,　shape　and　its　distribution,　the　K-median　clustering　method　is　used　to　select　a　more　appropriate　size　of　multi-scale　anchor　boxes　to　achieve　a　more　consistent　detection　boxes　for　the　log　ends;　(2)　The　detection　scales　of　YOLOv4-Tiny　network　are　increased　to　three　and　the　spatial　pyramid　pooling　(SPP)　module　is　added　to　enhance　the　ability　of　feature　extraction　for　small　targets,　such　as　small　log　end;　(3)　The　self-attention　mechanism　based　on　squeeze　and　excitation　(SE)　is　inserted　into　the　deep　structure　of　the　network,　which　automatically　determines　the　threshold　without　relevant　professional　knowledge　to　eliminate　the　high-dimensional　noise　that　may　result　in　error　gathering　of　the　centers　of　log　ends　and　predicted　boxes.　The　Preci-sion,　Recall　and　F1-score　of　YOLOv4-Tiny　in　the　experimental　test　set　are　91.87%,　94.91%　and　0.93　respectively,　while　the　three　indicators　are　improved　to　93.97%,　95.34%　and　0.95　respectively　by　using　the　proposed　model.　Moreover,　the　complete　intersection　over　union　(CIoU)　loss　of　our　model　is　2.46　and　is　reduced　by　51.48%　compared　to　the　YOLOv4-Tiny　with　5.07,　which　means　the　predicted　boxes　of　our　model　are　closer　to　the　target　bounding　boxes.　Consequently,　the　experiment　results　demonstrate　that　the　performance　of　the　proposed　approach　is　better　than　that　of　YOLOv4-Tiny　network.