diff --git a/internal/deepdoc/parser/pdf/inference/client.go b/internal/deepdoc/parser/pdf/inference/client.go
index c2b8fb95bf..451c2fdde1 100644
--- a/internal/deepdoc/parser/pdf/inference/client.go
+++ b/internal/deepdoc/parser/pdf/inference/client.go
@@ -76,6 +76,24 @@ type bboxesResponse struct {
 	BBoxes [][]float64 `json:"bboxes"`
 }
 
+// dlaGarbageLayouts mirrors Python LayoutRecognizer's garbage gate
+// (deepdoc/vision/layout_recognizer.py:97 and :379), which drops any region
+// whose type is in garbage_layouts=["footer","header","reference"] AND whose
+// confidence is below 0.4. We apply the same gate at the DLA source so every
+// consumer of Client.DLA (not just the table-annotation path, which re-applies
+// it downstream) sees the Python-aligned region set.
+//
+// Of the three types, only "reference" is reachable with the OSS default
+// 10-class DLA taxonomy (DefaultDLALabels has no footer/header classes), so in
+// practice this gate only fires on low-confidence references. footer/header are
+// included defensively to match Python's full garbage set for any deployment
+// whose DLA label taxonomy emits them.
+var dlaGarbageLayouts = map[string]bool{
+	string(pdf.LayoutTypeFooter):    true,
+	string(pdf.LayoutTypeHeader):    true,
+	string(pdf.LayoutTypeReference): true,
+}
+
 // DLA analyzes a full page image and returns labeled regions.
 func (c *Client) DLA(ctx context.Context, pageImage image.Image) ([]pdf.DLARegion, error) {
 	data, err := util.EncodePNG(pageImage)
@@ -96,6 +114,10 @@ func (c *Client) DLA(ctx context.Context, pageImage image.Image) ([]pdf.DLARegio
 		if clsID := int(b[5]); clsID >= 0 && clsID < len(labels) {
 			label = labels[clsID]
 		}
+		// Drop low-confidence garbage-layout regions (Python parity: 0.4 gate).
+		if dlaGarbageLayouts[label] && b[4] < 0.4 {
+			continue
+		}
 		regions = append(regions, pdf.DLARegion{
 			X0: b[0], Y0: b[1], X1: b[2], Y1: b[3],
 			Confidence: b[4],
diff --git a/internal/deepdoc/parser/pdf/inference/client_dla_garbage_test.go b/internal/deepdoc/parser/pdf/inference/client_dla_garbage_test.go
new file mode 100644
index 0000000000..199de8e5c4
--- /dev/null
+++ b/internal/deepdoc/parser/pdf/inference/client_dla_garbage_test.go
@@ -0,0 +1,97 @@
+package inference
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+// TestDeepDocHTTP_DLA_GarbageGate pins the Python-parity 0.4 garbage gate
+// (LayoutRecognizer.__call__, deepdoc/vision/layout_recognizer.py:97 and :379):
+// a region whose layout type is a garbage layout (footer/header/reference) and
+// whose confidence is strictly below 0.4 is dropped; everything else is kept.
+//
+// The OSS default 10-class DLA taxonomy only emits "reference" as a garbage
+// type, but footer/header are covered defensively (see dlaGarbageLayouts in
+// client.go). Each subtest drives a mock /predict/dla backend returning one
+// bbox and asserts the resulting region set.
+func TestDeepDocHTTP_DLA_GarbageGate(t *testing.T) {
+	newClient := func(t *testing.T, bboxes [][]float64) *Client {
+		t.Helper()
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path != "/predict/dla" {
+				t.Errorf("path = %q, want /predict/dla", r.URL.Path)
+			}
+			if err := json.NewEncoder(w).Encode(map[string]any{"bboxes": bboxes}); err != nil {
+				// Handler runs in its own goroutine; fail the test, don't panic.
+				t.Errorf("encode response: %v", err)
+			}
+		}))
+		t.Cleanup(srv.Close)
+		return mustNewDeepDocClient(t, srv.URL)
+	}
+
+	// bbox = [x0, y0, x1, y1, confidence, classId]; classId 2 = "reference".
+	const referenceClass = 2
+
+	t.Run("low_conf_reference_dropped", func(t *testing.T) {
+		client := newClient(t, [][]float64{
+			{50, 10, 500, 50, 0.30, referenceClass}, // reference, low confidence
+		})
+		regions, err := client.DLA(context.Background(), testImage())
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(regions) != 0 {
+			t.Fatalf("got %d regions, want 0 (low-confidence 'reference' must be dropped by the 0.4 garbage gate)", len(regions))
+		}
+	})
+
+	t.Run("high_conf_reference_kept", func(t *testing.T) {
+		client := newClient(t, [][]float64{
+			{50, 10, 500, 50, 0.90, referenceClass}, // reference, high confidence
+		})
+		regions, err := client.DLA(context.Background(), testImage())
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(regions) != 1 || regions[0].Label != "reference" {
+			t.Fatalf("got %v, want exactly one 'reference' region (conf >= 0.4 is kept)", regions)
+		}
+	})
+
+	t.Run("boundary_conf_0.4_kept", func(t *testing.T) {
+		// Gate is strict (< 0.4); confidence exactly 0.4 is the boundary and kept.
+		client := newClient(t, [][]float64{
+			{50, 10, 500, 50, 0.40, referenceClass}, // reference, exactly 0.4
+		})
+		regions, err := client.DLA(context.Background(), testImage())
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(regions) != 1 || regions[0].Label != "reference" {
+			t.Fatalf("got %v, want exactly one 'reference' region (conf == 0.4 is the gate boundary and kept)", regions)
+		}
+	})
+
+	t.Run("low_conf_garbage_and_text", func(t *testing.T) {
+		// A low-confidence reference is dropped while an unrelated text region
+		// is kept — Python keeps only the high-confidence non-garbage region.
+		client := newClient(t, [][]float64{
+			{50, 10, 500, 50, 0.30, referenceClass}, // reference, low confidence -> dropped
+			{50, 100, 500, 300, 0.90, 1},            // text, high confidence -> kept
+		})
+		regions, err := client.DLA(context.Background(), testImage())
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(regions) != 1 {
+			t.Fatalf("got %d regions, want 1 (low-confidence 'reference' dropped, 'text' kept)", len(regions))
+		}
+		if regions[0].Label != "text" {
+			t.Errorf("regions[0].Label = %q, want 'text'", regions[0].Label)
+		}
+	})
+}