From 59e378791c701c0f0d159ff2b5a5c7bf2f87e8b7 Mon Sep 17 00:00:00 2001
From: Nachtzuster <Nachtzuster@users.noreply.github.com>
Date: Sat, 27 Sep 2025 13:41:33 +0200
Subject: [PATCH] cleanup privacy filter, also make sure a human chunk their
 neighbour is also overwritten (#475)

---
 scripts/server.py | 96 +++++++++++++++++++++++++++++++----------------
 1 file changed, 63 insertions(+), 33 deletions(-)

diff --git a/scripts/server.py b/scripts/server.py
index ef16f34..b9a88a3 100644
--- a/scripts/server.py
+++ b/scripts/server.py
@@ -1,4 +1,3 @@
-import datetime
 import logging
 import math
 import operator
@@ -25,7 +24,7 @@ userDir = os.path.expanduser('~')
 INTERPRETER, M_INTERPRETER, INCLUDE_LIST, EXCLUDE_LIST = (None, None, None, None)
 PREDICTED_SPECIES_LIST = []
 WEEK = None
-model, priv_thresh, sf_thresh = (None, None, None)
+model, sf_thresh = (None, None)
 
 mdata, mdata_params = (None, None)
 
@@ -222,17 +221,7 @@ def predict(sample, sensitivity):
     # Sort by score
     p_sorted = sorted(p_labels.items(), key=operator.itemgetter(1), reverse=True)
 
-    human_cutoff = max(10, int(len(p_sorted) * priv_thresh / 100.0))
-
-    log.debug("DATABASE SIZE: %d", len(p_sorted))
-    log.debug("HUMAN-CUTOFF AT: %d", human_cutoff)
-
-    for i in range(min(10, len(p_sorted))):
-        if p_sorted[i][0] == 'Human_Human':
-            with open(userDir + '/BirdNET-Pi/HUMAN.txt', 'a') as rfile:
-                rfile.write(str(datetime.datetime.now()) + str(p_sorted[i]) + ' ' + str(human_cutoff) + '\n')
-
-    return p_sorted[:human_cutoff]
+    return p_sorted
 
 
 def analyzeAudioData(chunks, lat, lon, week, sens, overlap,):
@@ -240,7 +229,7 @@ def analyzeAudioData(chunks, lat, lon, week, sens, overlap,):
 
     sensitivity = max(0.5, min(1.0 - (sens - 1.0), 1.5))
 
-    detections = {}
+    detections = []
     start = time.time()
     log.info('ANALYZING AUDIO...')
 
@@ -252,35 +241,77 @@ def analyzeAudioData(chunks, lat, lon, week, sens, overlap,):
     mdata = get_metadata(lat, lon, week)
 
     # Parse every chunk
-    pred_start = 0.0
     for c in chunks:
-
         # Prepare as input signal
         sig = np.expand_dims(c, 0)
 
         # Make prediction
         p = predict([sig, mdata], sensitivity)
-#        print("PPPPP",p)
-        HUMAN_DETECTED = False
+        log.debug("PPPPP: %s", p)
+        detections.append(p)
 
-        # Catch if Human is recognized
-        for x in range(len(p)):
-            if "Human" in p[x][0]:
-                HUMAN_DETECTED = True
-
-        # Save result and timestamp
+    labeled = {}
+    pred_start = 0.0
+    for p in filter_humans(detections):
+        # Save timestamp and result
         pred_end = pred_start + 3.0
-
-        # If human detected set all detections to human to make sure voices are not saved
-        if HUMAN_DETECTED is True:
-            p = [('Human_Human', 0.0)] * 10
-
-        detections[str(pred_start) + ';' + str(pred_end)] = p
+        labeled[str(pred_start) + ';' + str(pred_end)] = p
 
         pred_start = pred_end - overlap
 
     log.info('DONE! Time %.2f SECONDS', time.time() - start)
-    return detections
+    return labeled
+
+
+def filter_humans(detections):
+    conf = get_settings()
+    priv_thresh = conf.getfloat('PRIVACY_THRESHOLD')
+    human_cutoff = max(10, int(len(detections[0]) * priv_thresh / 100.0))
+    log.debug("DATABASE SIZE: %d", len(detections[0]))
+    log.debug("HUMAN-CUTOFF AT: %d", human_cutoff)
+
+    censored_detections = []
+    for detection in detections:
+        p = detection[:human_cutoff]
+        human_detected = False
+        # Catch if Human is recognized in any of the predictions
+        for x in p:
+            if 'Human' in x[0]:
+                human_detected = True
+
+        # If human detected set detection to human to make sure voices are not saved
+        if human_detected is True:
+            p = [('Human_Human', 0.0)]
+        else:
+            p = p[:10]
+
+        censored_detections.append(p)
+
+    # now overwrite detections that have a human neighbour too
+    try:
+        extraction_length = conf.getint('EXTRACTION_LENGTH')
+    except ValueError:
+        extraction_length = 6
+    if extraction_length > 9:
+        log.warning("EXTRACTION_LENGTH is set to %d. Privacy filter might miss human sound, "
+                    "if you care about privacy, set EXTRACTION_LENGTH to below 9 or leave empty.", extraction_length)
+    human_neighbour_mask = [False] * len(censored_detections)
+    for i, detection in enumerate(censored_detections):
+        if i != 0:
+            if censored_detections[i - 1][0][0] == 'Human_Human':
+                human_neighbour_mask[i] = True
+        if i != len(censored_detections) - 1:
+            if censored_detections[i + 1][0][0] == 'Human_Human':
+                human_neighbour_mask[i] = True
+
+    clean_detections = []
+    for i, (has_human_neighbour, detection) in enumerate(zip(human_neighbour_mask, censored_detections)):
+        if has_human_neighbour and detection[0][0] != 'Human_Human':
+            log.debug('Overwriting detection %d %s - Has Human neighbour', i + 1, detection[0])
+            detection = [('Human_Human', 0.0)]
+        clean_detections.append(detection)
+
+    return clean_detections
 
 
 def get_metadata(lat, lon, week):
@@ -296,10 +327,9 @@ def get_metadata(lat, lon, week):
 
 def load_global_model():
     global INTERPRETER
-    global model, priv_thresh, sf_thresh
+    global model, sf_thresh
     conf = get_settings()
     model = conf['MODEL']
-    priv_thresh = conf.getfloat('PRIVACY_THRESHOLD')
     sf_thresh = conf.getfloat('SF_THRESH')
     INTERPRETER = loadModel()