Spaces:
Sleeping
Sleeping
Commit
·
b6c3e5f
1
Parent(s):
8d8e33d
feat: melhoria na interface com adição de gráficos e métricas de desempenho
Browse files
app.py
CHANGED
|
@@ -1,6 +1,10 @@
|
|
| 1 |
import cv2
|
| 2 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
| 3 |
import onnxruntime as ort
|
|
|
|
| 4 |
import gradio as gr
|
| 5 |
import os
|
| 6 |
from huggingface_hub import hf_hub_download
|
|
@@ -11,11 +15,12 @@ FILENAME = "tune/trial_10/weights/best.onnx"
|
|
| 11 |
MODEL_DIR = "model"
|
| 12 |
MODEL_PATH = os.path.join(MODEL_DIR, "model.onnx")
|
| 13 |
|
|
|
|
| 14 |
def download_model():
|
| 15 |
"""Download the model using Hugging Face Hub"""
|
| 16 |
# Ensure model directory exists
|
| 17 |
os.makedirs(MODEL_DIR, exist_ok=True)
|
| 18 |
-
|
| 19 |
try:
|
| 20 |
print(f"Downloading model from {REPO_ID}...")
|
| 21 |
# Download the model file from Hugging Face Hub
|
|
@@ -25,138 +30,179 @@ def download_model():
|
|
| 25 |
local_dir=MODEL_DIR,
|
| 26 |
local_dir_use_symlinks=False,
|
| 27 |
force_download=True,
|
| 28 |
-
cache_dir=None
|
| 29 |
)
|
| 30 |
-
|
| 31 |
# Move the file to the correct location if it's not there already
|
| 32 |
if os.path.exists(model_path) and model_path != MODEL_PATH:
|
| 33 |
os.rename(model_path, MODEL_PATH)
|
| 34 |
-
|
| 35 |
# Remove empty directories if they exist
|
| 36 |
empty_dir = os.path.join(MODEL_DIR, "tune")
|
| 37 |
if os.path.exists(empty_dir):
|
| 38 |
import shutil
|
|
|
|
| 39 |
shutil.rmtree(empty_dir)
|
| 40 |
-
|
| 41 |
print("Model downloaded successfully!")
|
| 42 |
return MODEL_PATH
|
| 43 |
-
|
| 44 |
except Exception as e:
|
| 45 |
-
print(f"Error downloading model: {
|
| 46 |
raise e
|
| 47 |
-
|
|
|
|
| 48 |
class SignatureDetector:
|
| 49 |
def __init__(self, model_path):
|
| 50 |
self.model_path = model_path
|
| 51 |
self.classes = ["signature"]
|
| 52 |
self.input_width = 640
|
| 53 |
self.input_height = 640
|
| 54 |
-
|
| 55 |
# Initialize ONNX Runtime session
|
| 56 |
-
self.session = ort.InferenceSession(
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
def preprocess(self, img):
|
| 59 |
# Convert PIL Image to cv2 format
|
| 60 |
img_cv2 = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
|
| 61 |
-
|
| 62 |
# Get image dimensions
|
| 63 |
self.img_height, self.img_width = img_cv2.shape[:2]
|
| 64 |
-
|
| 65 |
# Convert back to RGB for processing
|
| 66 |
img_rgb = cv2.cvtColor(img_cv2, cv2.COLOR_BGR2RGB)
|
| 67 |
-
|
| 68 |
# Resize
|
| 69 |
img_resized = cv2.resize(img_rgb, (self.input_width, self.input_height))
|
| 70 |
-
|
| 71 |
# Normalize and transpose
|
| 72 |
image_data = np.array(img_resized) / 255.0
|
| 73 |
image_data = np.transpose(image_data, (2, 0, 1))
|
| 74 |
image_data = np.expand_dims(image_data, axis=0).astype(np.float32)
|
| 75 |
-
|
| 76 |
return image_data, img_cv2
|
| 77 |
-
|
| 78 |
def draw_detections(self, img, box, score, class_id):
|
| 79 |
x1, y1, w, h = box
|
| 80 |
self.color_palette = np.random.uniform(0, 255, size=(len(self.classes), 3))
|
| 81 |
color = self.color_palette[class_id]
|
| 82 |
-
|
| 83 |
cv2.rectangle(img, (int(x1), int(y1)), (int(x1 + w), int(y1 + h)), color, 2)
|
| 84 |
-
|
| 85 |
label = f"{self.classes[class_id]}: {score:.2f}"
|
| 86 |
-
(label_width, label_height), _ = cv2.getTextSize(
|
| 87 |
-
|
|
|
|
|
|
|
| 88 |
label_x = x1
|
| 89 |
label_y = y1 - 10 if y1 - 10 > label_height else y1 + 10
|
| 90 |
-
|
| 91 |
cv2.rectangle(
|
| 92 |
img,
|
| 93 |
(int(label_x), int(label_y - label_height)),
|
| 94 |
(int(label_x + label_width), int(label_y + label_height)),
|
| 95 |
color,
|
| 96 |
-
cv2.FILLED
|
| 97 |
)
|
| 98 |
-
|
| 99 |
-
cv2.putText(
|
| 100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
def postprocess(self, input_image, output, conf_thres, iou_thres):
|
| 102 |
outputs = np.transpose(np.squeeze(output[0]))
|
| 103 |
rows = outputs.shape[0]
|
| 104 |
-
|
| 105 |
boxes = []
|
| 106 |
scores = []
|
| 107 |
class_ids = []
|
| 108 |
-
|
| 109 |
x_factor = self.img_width / self.input_width
|
| 110 |
y_factor = self.img_height / self.input_height
|
| 111 |
-
|
| 112 |
for i in range(rows):
|
| 113 |
classes_scores = outputs[i][4:]
|
| 114 |
max_score = np.amax(classes_scores)
|
| 115 |
-
|
| 116 |
if max_score >= conf_thres:
|
| 117 |
class_id = np.argmax(classes_scores)
|
| 118 |
x, y, w, h = outputs[i][0], outputs[i][1], outputs[i][2], outputs[i][3]
|
| 119 |
-
|
| 120 |
left = int((x - w / 2) * x_factor)
|
| 121 |
top = int((y - h / 2) * y_factor)
|
| 122 |
width = int(w * x_factor)
|
| 123 |
height = int(h * y_factor)
|
| 124 |
-
|
| 125 |
class_ids.append(class_id)
|
| 126 |
scores.append(max_score)
|
| 127 |
boxes.append([left, top, width, height])
|
| 128 |
-
|
| 129 |
indices = cv2.dnn.NMSBoxes(boxes, scores, conf_thres, iou_thres)
|
| 130 |
-
|
| 131 |
for i in indices:
|
| 132 |
box = boxes[i]
|
| 133 |
score = scores[i]
|
| 134 |
class_id = class_ids[i]
|
| 135 |
self.draw_detections(input_image, box, score, class_id)
|
| 136 |
-
|
| 137 |
return cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB)
|
| 138 |
-
|
| 139 |
def detect(self, image, conf_thres=0.25, iou_thres=0.5):
|
| 140 |
# Preprocess the image
|
| 141 |
img_data, original_image = self.preprocess(image)
|
| 142 |
-
|
| 143 |
# Run inference
|
|
|
|
| 144 |
outputs = self.session.run(None, {self.session.get_inputs()[0].name: img_data})
|
| 145 |
-
|
|
|
|
| 146 |
# Postprocess the results
|
| 147 |
output_image = self.postprocess(original_image, outputs, conf_thres, iou_thres)
|
| 148 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
return output_image
|
| 150 |
|
|
|
|
| 151 |
def create_gradio_interface():
|
| 152 |
# Download model if it doesn't exist
|
| 153 |
if not os.path.exists(MODEL_PATH):
|
| 154 |
download_model()
|
| 155 |
-
|
| 156 |
# Initialize the detector
|
| 157 |
detector = SignatureDetector(MODEL_PATH)
|
| 158 |
-
|
| 159 |
-
|
| 160 |
css = """
|
| 161 |
.custom-button {
|
| 162 |
background-color: #b0ffb8 !important;
|
|
@@ -165,24 +211,125 @@ def create_gradio_interface():
|
|
| 165 |
.custom-button:hover {
|
| 166 |
background-color: #b0ffb8b3 !important;
|
| 167 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 168 |
"""
|
| 169 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
with gr.Blocks(
|
| 171 |
-
theme
|
| 172 |
-
primary_hue="indigo",
|
| 173 |
-
secondary_hue="gray",
|
| 174 |
-
neutral_hue="gray"
|
| 175 |
),
|
| 176 |
-
css=css
|
| 177 |
) as iface:
|
| 178 |
gr.Markdown(
|
| 179 |
"""
|
| 180 |
# Tech4Humans - Detector de Assinaturas
|
| 181 |
|
| 182 |
Este sistema utiliza o modelo [**YOLOv8s**](https://huggingface.co/tech4humans/yolov8s-signature-detector), especialmente ajustado para a detecção de assinaturas manuscritas em imagens de documentos.
|
| 183 |
-
|
| 184 |
-
mecanismos de pré-processamento e aumento de dados para garantir alta precisão e generalização.
|
| 185 |
-
|
| 186 |
Com este detector, é possível identificar assinaturas em documentos digitais com elevada precisão em tempo real, sendo ideal para
|
| 187 |
aplicações que envolvem validação, organização e processamento de documentos.
|
| 188 |
|
|
@@ -190,84 +337,100 @@ def create_gradio_interface():
|
|
| 190 |
"""
|
| 191 |
)
|
| 192 |
|
| 193 |
-
with gr.Row():
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
|
|
|
|
|
|
|
|
|
| 198 |
clear_btn = gr.ClearButton([input_image], value="Limpar")
|
| 199 |
submit_btn = gr.Button("Detectar", elem_classes="custom-button")
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
)
|
| 217 |
-
|
| 218 |
-
output_image = gr.Image(label="Resultados da Detecção") # Em outra coluna
|
| 219 |
-
|
| 220 |
-
clear_btn.add(output_image)
|
| 221 |
-
|
| 222 |
-
gr.Examples(
|
| 223 |
-
examples=[
|
| 224 |
-
["assets/images/example_{i}.jpg".format(i=i)] for i in range(0, len(os.listdir(os.path.join("assets", "images"))))
|
| 225 |
-
],
|
| 226 |
-
inputs=input_image,
|
| 227 |
-
outputs=output_image,
|
| 228 |
-
fn=detector.detect,
|
| 229 |
-
label="Exemplos",
|
| 230 |
-
cache_examples=True,
|
| 231 |
-
cache_mode='lazy'
|
| 232 |
-
)
|
| 233 |
|
|
|
|
|
|
|
| 234 |
|
| 235 |
-
|
| 236 |
-
fn=detector.detect,
|
| 237 |
-
inputs=[input_image, confidence_threshold, iou_threshold],
|
| 238 |
-
outputs=output_image,
|
| 239 |
-
)
|
| 240 |
-
|
| 241 |
-
gr.Markdown(
|
| 242 |
-
"""
|
| 243 |
-
---
|
| 244 |
-
## Sobre o Modelo e Resultados
|
| 245 |
|
| 246 |
-
|
|
|
|
|
|
|
|
|
|
| 247 |
|
| 248 |
-
|
| 249 |
-
- **Precisão (Precision):** 94,74%
|
| 250 |
-
- **Revocação (Recall):** 89,72%
|
| 251 |
-
- **mAP@50:** 94,50%
|
| 252 |
-
- **mAP@50-95:** 67,35%
|
| 253 |
-
- **Tempo de Inferência (CPU):** 171,56 ms
|
| 254 |
|
| 255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
|
| 257 |
-
|
| 258 |
|
| 259 |
-
|
| 260 |
-
"""
|
| 261 |
-
)
|
| 262 |
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 267 |
)
|
| 268 |
-
|
| 269 |
return iface
|
| 270 |
|
|
|
|
| 271 |
if __name__ == "__main__":
|
| 272 |
iface = create_gradio_interface()
|
| 273 |
-
iface.launch()
|
|
|
|
| 1 |
import cv2
|
| 2 |
import numpy as np
|
| 3 |
+
import pandas as pd
|
| 4 |
+
import time
|
| 5 |
+
import matplotlib.pyplot as plt
|
| 6 |
import onnxruntime as ort
|
| 7 |
+
from collections import deque
|
| 8 |
import gradio as gr
|
| 9 |
import os
|
| 10 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 15 |
MODEL_DIR = "model"
|
| 16 |
MODEL_PATH = os.path.join(MODEL_DIR, "model.onnx")
|
| 17 |
|
| 18 |
+
|
| 19 |
def download_model():
|
| 20 |
"""Download the model using Hugging Face Hub"""
|
| 21 |
# Ensure model directory exists
|
| 22 |
os.makedirs(MODEL_DIR, exist_ok=True)
|
| 23 |
+
|
| 24 |
try:
|
| 25 |
print(f"Downloading model from {REPO_ID}...")
|
| 26 |
# Download the model file from Hugging Face Hub
|
|
|
|
| 30 |
local_dir=MODEL_DIR,
|
| 31 |
local_dir_use_symlinks=False,
|
| 32 |
force_download=True,
|
| 33 |
+
cache_dir=None,
|
| 34 |
)
|
| 35 |
+
|
| 36 |
# Move the file to the correct location if it's not there already
|
| 37 |
if os.path.exists(model_path) and model_path != MODEL_PATH:
|
| 38 |
os.rename(model_path, MODEL_PATH)
|
| 39 |
+
|
| 40 |
# Remove empty directories if they exist
|
| 41 |
empty_dir = os.path.join(MODEL_DIR, "tune")
|
| 42 |
if os.path.exists(empty_dir):
|
| 43 |
import shutil
|
| 44 |
+
|
| 45 |
shutil.rmtree(empty_dir)
|
| 46 |
+
|
| 47 |
print("Model downloaded successfully!")
|
| 48 |
return MODEL_PATH
|
| 49 |
+
|
| 50 |
except Exception as e:
|
| 51 |
+
print(f"Error downloading model: {e}")
|
| 52 |
raise e
|
| 53 |
+
|
| 54 |
+
|
| 55 |
class SignatureDetector:
|
| 56 |
def __init__(self, model_path):
|
| 57 |
self.model_path = model_path
|
| 58 |
self.classes = ["signature"]
|
| 59 |
self.input_width = 640
|
| 60 |
self.input_height = 640
|
| 61 |
+
|
| 62 |
# Initialize ONNX Runtime session
|
| 63 |
+
self.session = ort.InferenceSession(
|
| 64 |
+
MODEL_PATH, providers=["CPUExecutionProvider"]
|
| 65 |
+
)
|
| 66 |
+
|
| 67 |
+
# Initialize metrics tracking
|
| 68 |
+
self.inference_times = deque(maxlen=50) # Store last 50 inference times
|
| 69 |
+
self.total_inferences = 0
|
| 70 |
+
self.avg_inference_time = 0
|
| 71 |
+
|
| 72 |
+
def update_metrics(self, inference_time):
|
| 73 |
+
self.inference_times.append(inference_time)
|
| 74 |
+
self.total_inferences += 1
|
| 75 |
+
self.avg_inference_time = sum(self.inference_times) / len(self.inference_times)
|
| 76 |
+
|
| 77 |
+
def get_metrics(self):
|
| 78 |
+
return {
|
| 79 |
+
"times": list(self.inference_times),
|
| 80 |
+
"total_inferences": self.total_inferences,
|
| 81 |
+
"avg_time": self.avg_inference_time,
|
| 82 |
+
}
|
| 83 |
+
|
| 84 |
def preprocess(self, img):
|
| 85 |
# Convert PIL Image to cv2 format
|
| 86 |
img_cv2 = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
|
| 87 |
+
|
| 88 |
# Get image dimensions
|
| 89 |
self.img_height, self.img_width = img_cv2.shape[:2]
|
| 90 |
+
|
| 91 |
# Convert back to RGB for processing
|
| 92 |
img_rgb = cv2.cvtColor(img_cv2, cv2.COLOR_BGR2RGB)
|
| 93 |
+
|
| 94 |
# Resize
|
| 95 |
img_resized = cv2.resize(img_rgb, (self.input_width, self.input_height))
|
| 96 |
+
|
| 97 |
# Normalize and transpose
|
| 98 |
image_data = np.array(img_resized) / 255.0
|
| 99 |
image_data = np.transpose(image_data, (2, 0, 1))
|
| 100 |
image_data = np.expand_dims(image_data, axis=0).astype(np.float32)
|
| 101 |
+
|
| 102 |
return image_data, img_cv2
|
| 103 |
+
|
| 104 |
def draw_detections(self, img, box, score, class_id):
|
| 105 |
x1, y1, w, h = box
|
| 106 |
self.color_palette = np.random.uniform(0, 255, size=(len(self.classes), 3))
|
| 107 |
color = self.color_palette[class_id]
|
| 108 |
+
|
| 109 |
cv2.rectangle(img, (int(x1), int(y1)), (int(x1 + w), int(y1 + h)), color, 2)
|
| 110 |
+
|
| 111 |
label = f"{self.classes[class_id]}: {score:.2f}"
|
| 112 |
+
(label_width, label_height), _ = cv2.getTextSize(
|
| 113 |
+
label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1
|
| 114 |
+
)
|
| 115 |
+
|
| 116 |
label_x = x1
|
| 117 |
label_y = y1 - 10 if y1 - 10 > label_height else y1 + 10
|
| 118 |
+
|
| 119 |
cv2.rectangle(
|
| 120 |
img,
|
| 121 |
(int(label_x), int(label_y - label_height)),
|
| 122 |
(int(label_x + label_width), int(label_y + label_height)),
|
| 123 |
color,
|
| 124 |
+
cv2.FILLED,
|
| 125 |
)
|
| 126 |
+
|
| 127 |
+
cv2.putText(
|
| 128 |
+
img,
|
| 129 |
+
label,
|
| 130 |
+
(int(label_x), int(label_y)),
|
| 131 |
+
cv2.FONT_HERSHEY_SIMPLEX,
|
| 132 |
+
0.5,
|
| 133 |
+
(0, 0, 0),
|
| 134 |
+
1,
|
| 135 |
+
cv2.LINE_AA,
|
| 136 |
+
)
|
| 137 |
+
|
| 138 |
def postprocess(self, input_image, output, conf_thres, iou_thres):
|
| 139 |
outputs = np.transpose(np.squeeze(output[0]))
|
| 140 |
rows = outputs.shape[0]
|
| 141 |
+
|
| 142 |
boxes = []
|
| 143 |
scores = []
|
| 144 |
class_ids = []
|
| 145 |
+
|
| 146 |
x_factor = self.img_width / self.input_width
|
| 147 |
y_factor = self.img_height / self.input_height
|
| 148 |
+
|
| 149 |
for i in range(rows):
|
| 150 |
classes_scores = outputs[i][4:]
|
| 151 |
max_score = np.amax(classes_scores)
|
| 152 |
+
|
| 153 |
if max_score >= conf_thres:
|
| 154 |
class_id = np.argmax(classes_scores)
|
| 155 |
x, y, w, h = outputs[i][0], outputs[i][1], outputs[i][2], outputs[i][3]
|
| 156 |
+
|
| 157 |
left = int((x - w / 2) * x_factor)
|
| 158 |
top = int((y - h / 2) * y_factor)
|
| 159 |
width = int(w * x_factor)
|
| 160 |
height = int(h * y_factor)
|
| 161 |
+
|
| 162 |
class_ids.append(class_id)
|
| 163 |
scores.append(max_score)
|
| 164 |
boxes.append([left, top, width, height])
|
| 165 |
+
|
| 166 |
indices = cv2.dnn.NMSBoxes(boxes, scores, conf_thres, iou_thres)
|
| 167 |
+
|
| 168 |
for i in indices:
|
| 169 |
box = boxes[i]
|
| 170 |
score = scores[i]
|
| 171 |
class_id = class_ids[i]
|
| 172 |
self.draw_detections(input_image, box, score, class_id)
|
| 173 |
+
|
| 174 |
return cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB)
|
| 175 |
+
|
| 176 |
def detect(self, image, conf_thres=0.25, iou_thres=0.5):
|
| 177 |
# Preprocess the image
|
| 178 |
img_data, original_image = self.preprocess(image)
|
| 179 |
+
|
| 180 |
# Run inference
|
| 181 |
+
start_time = time.time()
|
| 182 |
outputs = self.session.run(None, {self.session.get_inputs()[0].name: img_data})
|
| 183 |
+
inference_time = (time.time() - start_time) * 1000 # Convert to milliseconds
|
| 184 |
+
|
| 185 |
# Postprocess the results
|
| 186 |
output_image = self.postprocess(original_image, outputs, conf_thres, iou_thres)
|
| 187 |
+
|
| 188 |
+
self.update_metrics(inference_time)
|
| 189 |
+
|
| 190 |
+
return output_image, self.get_metrics()
|
| 191 |
+
|
| 192 |
+
def detect_example(self, image, conf_thres=0.25, iou_thres=0.5):
|
| 193 |
+
"""Wrapper method for examples that returns only the image"""
|
| 194 |
+
output_image, _ = self.detect(image, conf_thres, iou_thres)
|
| 195 |
return output_image
|
| 196 |
|
| 197 |
+
|
| 198 |
def create_gradio_interface():
|
| 199 |
# Download model if it doesn't exist
|
| 200 |
if not os.path.exists(MODEL_PATH):
|
| 201 |
download_model()
|
| 202 |
+
|
| 203 |
# Initialize the detector
|
| 204 |
detector = SignatureDetector(MODEL_PATH)
|
| 205 |
+
|
|
|
|
| 206 |
css = """
|
| 207 |
.custom-button {
|
| 208 |
background-color: #b0ffb8 !important;
|
|
|
|
| 211 |
.custom-button:hover {
|
| 212 |
background-color: #b0ffb8b3 !important;
|
| 213 |
}
|
| 214 |
+
.container {
|
| 215 |
+
max-width: 1200px !important;
|
| 216 |
+
margin: auto !important;
|
| 217 |
+
}
|
| 218 |
+
.main-container {
|
| 219 |
+
gap: 20px !important;
|
| 220 |
+
}
|
| 221 |
+
.metrics-container {
|
| 222 |
+
padding: 1.5rem !important;
|
| 223 |
+
border-radius: 0.75rem !important;
|
| 224 |
+
background-color: #1f2937 !important;
|
| 225 |
+
margin: 1rem 0 !important;
|
| 226 |
+
box-shadow: 0 1px 3px rgba(0, 0, 0, 0.1) !important;
|
| 227 |
+
}
|
| 228 |
+
.metrics-title {
|
| 229 |
+
font-size: 1.25rem !important;
|
| 230 |
+
font-weight: 600 !important;
|
| 231 |
+
color: #1f2937 !important;
|
| 232 |
+
margin-bottom: 1rem !important;
|
| 233 |
+
}
|
| 234 |
"""
|
| 235 |
|
| 236 |
+
def process_image(image, conf_thres, iou_thres):
|
| 237 |
+
if image is None:
|
| 238 |
+
return None, None, None, None
|
| 239 |
+
|
| 240 |
+
output_image, metrics = detector.detect(image, conf_thres, iou_thres)
|
| 241 |
+
|
| 242 |
+
# Create plots data
|
| 243 |
+
hist_data = pd.DataFrame({"Tempo (ms)": metrics["times"]})
|
| 244 |
+
line_data = pd.DataFrame(
|
| 245 |
+
{
|
| 246 |
+
"Inferência": range(len(metrics["times"])),
|
| 247 |
+
"Tempo (ms)": metrics["times"],
|
| 248 |
+
"Média": [metrics["avg_time"]] * len(metrics["times"]),
|
| 249 |
+
}
|
| 250 |
+
)
|
| 251 |
+
|
| 252 |
+
# Limpar figuras existentes
|
| 253 |
+
plt.close("all")
|
| 254 |
+
|
| 255 |
+
# Configuração do estilo dos plots
|
| 256 |
+
plt.style.use("dark_background")
|
| 257 |
+
|
| 258 |
+
# Criar figura do histograma
|
| 259 |
+
hist_fig, hist_ax = plt.subplots(figsize=(8, 4), facecolor="#f0f0f5")
|
| 260 |
+
hist_ax.set_facecolor("#f0f0f5")
|
| 261 |
+
hist_data.hist(
|
| 262 |
+
bins=20, ax=hist_ax, color="#4F46E5", alpha=0.7, edgecolor="white"
|
| 263 |
+
)
|
| 264 |
+
hist_ax.set_title(
|
| 265 |
+
"Distribuição dos Tempos de Inferência",
|
| 266 |
+
pad=15,
|
| 267 |
+
fontsize=12,
|
| 268 |
+
color="#1f2937",
|
| 269 |
+
)
|
| 270 |
+
hist_ax.set_xlabel("Tempo (ms)", color="#374151")
|
| 271 |
+
hist_ax.set_ylabel("Frequência", color="#374151")
|
| 272 |
+
hist_ax.tick_params(colors="#4b5563")
|
| 273 |
+
hist_ax.grid(True, linestyle="--", alpha=0.3)
|
| 274 |
+
|
| 275 |
+
# Criar figura do gráfico de linha
|
| 276 |
+
line_fig, line_ax = plt.subplots(figsize=(8, 4), facecolor="#f0f0f5")
|
| 277 |
+
line_ax.set_facecolor("#f0f0f5")
|
| 278 |
+
line_data.plot(
|
| 279 |
+
x="Inferência",
|
| 280 |
+
y="Tempo (ms)",
|
| 281 |
+
ax=line_ax,
|
| 282 |
+
color="#4F46E5",
|
| 283 |
+
alpha=0.7,
|
| 284 |
+
label="Tempo",
|
| 285 |
+
)
|
| 286 |
+
line_data.plot(
|
| 287 |
+
x="Inferência",
|
| 288 |
+
y="Média",
|
| 289 |
+
ax=line_ax,
|
| 290 |
+
color="#DC2626",
|
| 291 |
+
linestyle="--",
|
| 292 |
+
label="Média",
|
| 293 |
+
)
|
| 294 |
+
line_ax.set_title(
|
| 295 |
+
"Tempo de Inferência por Execução", pad=15, fontsize=12, color="#1f2937"
|
| 296 |
+
)
|
| 297 |
+
line_ax.set_xlabel("Número da Inferência", color="#374151")
|
| 298 |
+
line_ax.set_ylabel("Tempo (ms)", color="#374151")
|
| 299 |
+
line_ax.tick_params(colors="#4b5563")
|
| 300 |
+
line_ax.grid(True, linestyle="--", alpha=0.3)
|
| 301 |
+
line_ax.legend(frameon=True, facecolor="#f0f0f5", edgecolor="none")
|
| 302 |
+
|
| 303 |
+
# Ajustar layout
|
| 304 |
+
hist_fig.tight_layout()
|
| 305 |
+
line_fig.tight_layout()
|
| 306 |
+
|
| 307 |
+
# Fechar as figuras para liberar memória
|
| 308 |
+
plt.close(hist_fig)
|
| 309 |
+
plt.close(line_fig)
|
| 310 |
+
|
| 311 |
+
return (
|
| 312 |
+
output_image,
|
| 313 |
+
gr.update(
|
| 314 |
+
value=f"Total de Inferências: {metrics['total_inferences']}",
|
| 315 |
+
container=True,
|
| 316 |
+
),
|
| 317 |
+
hist_fig,
|
| 318 |
+
line_fig,
|
| 319 |
+
)
|
| 320 |
+
|
| 321 |
with gr.Blocks(
|
| 322 |
+
theme=gr.themes.Soft(
|
| 323 |
+
primary_hue="indigo", secondary_hue="gray", neutral_hue="gray"
|
|
|
|
|
|
|
| 324 |
),
|
| 325 |
+
css=css,
|
| 326 |
) as iface:
|
| 327 |
gr.Markdown(
|
| 328 |
"""
|
| 329 |
# Tech4Humans - Detector de Assinaturas
|
| 330 |
|
| 331 |
Este sistema utiliza o modelo [**YOLOv8s**](https://huggingface.co/tech4humans/yolov8s-signature-detector), especialmente ajustado para a detecção de assinaturas manuscritas em imagens de documentos.
|
| 332 |
+
|
|
|
|
|
|
|
| 333 |
Com este detector, é possível identificar assinaturas em documentos digitais com elevada precisão em tempo real, sendo ideal para
|
| 334 |
aplicações que envolvem validação, organização e processamento de documentos.
|
| 335 |
|
|
|
|
| 337 |
"""
|
| 338 |
)
|
| 339 |
|
| 340 |
+
with gr.Row(equal_height=True, elem_classes="main-container"):
|
| 341 |
+
# Coluna da esquerda para controles e informações
|
| 342 |
+
with gr.Column(scale=1):
|
| 343 |
+
input_image = gr.Image(
|
| 344 |
+
label="Faça o upload do seu documento", type="pil"
|
| 345 |
+
)
|
| 346 |
+
|
| 347 |
+
with gr.Row():
|
| 348 |
clear_btn = gr.ClearButton([input_image], value="Limpar")
|
| 349 |
submit_btn = gr.Button("Detectar", elem_classes="custom-button")
|
| 350 |
+
|
| 351 |
+
with gr.Group():
|
| 352 |
+
confidence_threshold = gr.Slider(
|
| 353 |
+
minimum=0.0,
|
| 354 |
+
maximum=1.0,
|
| 355 |
+
value=0.25,
|
| 356 |
+
step=0.05,
|
| 357 |
+
label="Limiar de Confiança",
|
| 358 |
+
info="Ajuste a pontuação mínima de confiança necessária para detecção.",
|
| 359 |
+
)
|
| 360 |
+
iou_threshold = gr.Slider(
|
| 361 |
+
minimum=0.0,
|
| 362 |
+
maximum=1.0,
|
| 363 |
+
value=0.5,
|
| 364 |
+
step=0.05,
|
| 365 |
+
label="Limiar de IoU",
|
| 366 |
+
info="Ajuste o limiar de Interseção sobre União para Non Maximum Suppression (NMS).",
|
| 367 |
+
)
|
| 368 |
+
|
| 369 |
+
with gr.Column(scale=1):
|
| 370 |
+
output_image = gr.Image(label="Resultados da Detecção")
|
| 371 |
+
|
| 372 |
+
with gr.Accordion("Exemplos", open=True):
|
| 373 |
+
gr.Examples(
|
| 374 |
+
examples=[
|
| 375 |
+
["assets/images/example_{i}.jpg".format(i=i)]
|
| 376 |
+
for i in range(
|
| 377 |
+
0, len(os.listdir(os.path.join("assets", "images")))
|
| 378 |
+
)
|
| 379 |
+
],
|
| 380 |
+
inputs=input_image,
|
| 381 |
+
outputs=output_image,
|
| 382 |
+
fn=detector.detect_example,
|
| 383 |
+
cache_examples=True,
|
| 384 |
+
cache_mode="lazy",
|
| 385 |
+
)
|
| 386 |
+
|
| 387 |
+
with gr.Row(elem_classes="metrics-container"):
|
| 388 |
+
with gr.Column(scale=1):
|
| 389 |
+
total_inferences = gr.Textbox(
|
| 390 |
+
label="Total de Inferências", show_copy_button=True, container=True
|
| 391 |
)
|
| 392 |
+
hist_plot = gr.Plot(label="Distribuição dos Tempos", container=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 393 |
|
| 394 |
+
with gr.Column(scale=1):
|
| 395 |
+
line_plot = gr.Plot(label="Histórico de Tempos", container=True)
|
| 396 |
|
| 397 |
+
with gr.Row(elem_classes="container"):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 398 |
|
| 399 |
+
gr.Markdown(
|
| 400 |
+
"""
|
| 401 |
+
---
|
| 402 |
+
## Sobre o Projeto
|
| 403 |
|
| 404 |
+
Este projeto utiliza o modelo YOLOv8s ajustado para detecção de assinaturas manuscritas em imagens de documentos. Ele foi treinado com dados provenientes dos conjuntos [Tobacco800](https://paperswithcode.com/dataset/tobacco-800) e [signatures-xc8up](https://universe.roboflow.com/roboflow-100/signatures-xc8up), passando por processos de pré-processamento e aumentação de dados.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 405 |
|
| 406 |
+
### Principais Métricas:
|
| 407 |
+
- **Precisão (Precision):** 94,74%
|
| 408 |
+
- **Revocação (Recall):** 89,72%
|
| 409 |
+
- **mAP@50:** 94,50%
|
| 410 |
+
- **mAP@50-95:** 67,35%
|
| 411 |
+
- **Tempo de Inferência (CPU):** 171,56 ms
|
| 412 |
|
| 413 |
+
O processo completo de treinamento, ajuste de hiperparâmetros, e avaliação do modelo pode ser consultado em detalhes no repositório abaixo.
|
| 414 |
|
| 415 |
+
[Leia o README completo no Hugging Face Models](https://huggingface.co/tech4humans/yolov8s-signature-detector)
|
|
|
|
|
|
|
| 416 |
|
| 417 |
+
---
|
| 418 |
+
|
| 419 |
+
**Desenvolvido por [Tech4Humans](https://www.tech4h.com.br/)** | **Modelo:** [YOLOv8s](https://huggingface.co/tech4humans/yolov8s-signature-detector) | **Datasets:** [Tobacco800](https://paperswithcode.com/dataset/tobacco-800), [signatures-xc8up](https://universe.roboflow.com/roboflow-100/signatures-xc8up)
|
| 420 |
+
"""
|
| 421 |
+
)
|
| 422 |
+
|
| 423 |
+
clear_btn.add([output_image, total_inferences, hist_plot, line_plot])
|
| 424 |
+
|
| 425 |
+
submit_btn.click(
|
| 426 |
+
fn=process_image,
|
| 427 |
+
inputs=[input_image, confidence_threshold, iou_threshold],
|
| 428 |
+
outputs=[output_image, total_inferences, hist_plot, line_plot],
|
| 429 |
)
|
| 430 |
+
|
| 431 |
return iface
|
| 432 |
|
| 433 |
+
|
| 434 |
if __name__ == "__main__":
|
| 435 |
iface = create_gradio_interface()
|
| 436 |
+
iface.launch()
|