公招结果识别：mobilenet保留分类器前三层；先通过职业分类，再用knn分类

mower/solvers/recruit.py

@@ -18,10 +18,10 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
 
 from copy import copy
 from datetime import timedelta
+from functools import cache
 from itertools import combinations
 
 import cv2
-import numpy as np
 
 from mower.data import recruit_agent
 from mower.static import recruit, recruit_result_knn
@@ -32,7 +32,6 @@ from mower.utils.log import logger
 from mower.utils.path import get_path
 from mower.utils.scene import Scene
 from mower.utils.solver import BaseSolver
-from mower.utils.typealias import Image
 from mower.utils.vector import sa, va
 
 
@@ -91,8 +90,6 @@ class RecruitSolver(BaseSolver):
     solver_max_duration = timedelta(minutes=3)
 
     def run(self):
-        net_path = get_path("@install/mower/static/mobilenet_v3_small_features.onnx")
-        self.mobilenet = cv2.dnn.readNetFromONNX(str(net_path))
         self.index_known: bool = False
         self.slot_index: int = 0
         self.info = {}
@@ -235,18 +232,30 @@ class RecruitSolver(BaseSolver):
         self.ctap((x, y), 2)
         return True
 
+    @property
+    @cache
+    def mobilenet(self):
+        net_name = "mobilenet_v3_small_feature_extractor.onnx"
+        net_path = get_path(f"@install/mower/static/{net_name}")
+        return cv2.dnn.readNetFromONNX(str(net_path))
+
     def recruit_result(self) -> str:
-        img: Image = cropimg(config.recog.img, ((800, 100), (1400, 700)))
+        for profession, knn_classifier in recruit_result_knn().items():
-        mean = np.array([0.485, 0.456, 0.406], dtype=np.float32) * 255.0
+            if config.recog.find(f"recruit/profession/{profession}"):
-        std = np.array([0.229, 0.224, 0.225], dtype=np.float32) * 255.0
+                img = cropimg(config.recog.img, ((800, 100), (1300, 600)))
-        img = (img.astype(np.float32) - mean) / std
+                blob = cv2.dnn.blobFromImage(
-        img = np.transpose(img, (2, 0, 1))
+                    img,
-        img = np.expand_dims(img, axis=0).astype(np.float32)
+                    scalefactor=1 / 255,
-        self.mobilenet.setInput(img, "input")
+                    size=(224, 224),
-        features = self.mobilenet.forward("output")
+                    mean=(0, 0, 0),
-        result = recruit_result_knn().predict(features)
+                    swapRB=False,
-        logger.debug(result)
+                    crop=False,
-        return result
+                )
+                self.mobilenet.setInput(blob)
+                features = self.mobilenet.forward()
+                result = str(knn_classifier.predict(features)[0])
+                logger.debug(result)
+                return result
 
     def transition(self):
         if len(self.info) == 4:

mower/utils/recognize/data.py

@@ -408,14 +408,14 @@ template_matching = {
     "recruit/agent_token_first": ((1700, 760), (1920, 810)),
     "recruit/begin_recruit": None,
     "recruit/job_requirements": None,
-    "recruit/profession/CASTER": None,
+    "recruit/profession/CASTER": ((700, 720), (1040, 890)),
-    "recruit/profession/MEDIC": None,
+    "recruit/profession/MEDIC": ((700, 720), (1040, 890)),
-    "recruit/profession/PIONEER": None,
+    "recruit/profession/PIONEER": ((700, 720), (1040, 890)),
-    "recruit/profession/SNIPER": None,
+    "recruit/profession/SNIPER": ((700, 720), (1040, 890)),
-    "recruit/profession/SPECIAL": None,
+    "recruit/profession/SPECIAL": ((700, 720), (1040, 890)),
-    "recruit/profession/SUPPORT": None,
+    "recruit/profession/SUPPORT": ((700, 720), (1040, 890)),
-    "recruit/profession/TANK": None,
+    "recruit/profession/TANK": ((700, 720), (1040, 890)),
-    "recruit/profession/WARRIOR": None,
+    "recruit/profession/WARRIOR": ((700, 720), (1040, 890)),
     "recruit/recruit_done": None,
     "recruit/recruit_lock": None,
     "recruit/refresh_comfirm": (1237, 714),