XiaoBai1221 commited on
Commit
86d38f2
·
1 Parent(s): ab4f6fb

簡化 Gradio 介面設計 - 修復 JSON schema 錯誤和移除 share=True

Browse files
Files changed (1) hide show
  1. app.py +0 -187
app.py DELETED
@@ -1,187 +0,0 @@
1
- import os
2
- import cv2
3
- import numpy as np
4
- import gradio as gr
5
- from realtime_sign_prediction import RealtimeSignPredictor
6
-
7
- class GradioSignPredictor:
8
- def __init__(self):
9
- """初始化手語辨識預測器"""
10
- self.predictor = RealtimeSignPredictor(
11
- model_path="tsflow/models/best_model.pt",
12
- config_path="tsflow/results/test_results.json",
13
- sequence_length=50,
14
- use_segmentation=True
15
- )
16
- print("✅ 手語辨識系統初始化完成!")
17
-
18
- def process_frame(self, frame):
19
- """處理單一畫面並返回結果"""
20
- if frame is None:
21
- return frame, "等待攝像頭輸入..."
22
-
23
- try:
24
- # 處理畫面
25
- results, keypoints, flow_features = self.predictor.process_frame(frame)
26
-
27
- # 繪製關鍵點
28
- annotated_frame = self.predictor.draw_landmarks(frame.copy(), results)
29
-
30
- # 獲取預測結果
31
- top_predictions = self.predictor.get_top_predictions(top_k=3)
32
-
33
- # 格式化預測結果
34
- if top_predictions:
35
- prediction_text = "🎯 即時預測結果:\n\n"
36
- for i, (label, confidence) in enumerate(top_predictions, 1):
37
- confidence_bar = "█" * int(confidence * 20)
38
- prediction_text += f"{i}. {label}: {confidence:.2%} {confidence_bar}\n"
39
-
40
- # 添加序列資訊
41
- prediction_text += f"\n📊 序列進度: {len(self.predictor.keypoint_sequence)}/{self.predictor.sequence_length}"
42
- else:
43
- prediction_text = "📡 正在收集動作序列...\n請在攝像頭前做手語動作"
44
-
45
- return annotated_frame, prediction_text
46
-
47
- except Exception as e:
48
- return frame, f"❌ 處理錯誤: {str(e)}"
49
-
50
- def clear_predictions(self):
51
- """清除預測序列"""
52
- self.predictor.keypoint_sequence.clear()
53
- self.predictor.flow_sequence.clear()
54
- return "✅ 已清除預測序列,請重新開始"
55
-
56
- # 初始化全域預測器
57
- print("🚀 正在初始化手語辨識系統...")
58
- global_predictor = GradioSignPredictor()
59
-
60
- def process_video_frame(frame):
61
- """處理視訊畫面的包裝函數"""
62
- return global_predictor.process_frame(frame)
63
-
64
- def clear_predictions():
65
- """清除預測的包裝函數"""
66
- return global_predictor.clear_predictions()
67
-
68
- # 創建 Gradio 介面 - 修復 Gradio 4.44.x 兼容性
69
- with gr.Blocks(
70
- title="SignView2.0 - 手語辨識系統",
71
- theme=gr.themes.Soft(),
72
- css="""
73
- .container { max-width: 1200px; margin: auto; }
74
- .header { text-align: center; margin-bottom: 2rem; }
75
- .prediction-box {
76
- background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
77
- color: white;
78
- padding: 1rem;
79
- border-radius: 10px;
80
- font-family: monospace;
81
- }
82
- """
83
- ) as demo:
84
-
85
- gr.Markdown("""
86
- # 🤟 SignView2.0 - 手語辨識系統
87
-
88
- **支援34種手語詞彙的即時辨識系統,準確率達94.25%**
89
-
90
- ## 📋 支援詞彙
91
- `again`, `all`, `apple`, `bad`, `bathroom`, `beautiful`, `bird`, `black`, `blue`, `book`,
92
- `bored`, `boy`, `brother`, `brown`, `but`, `computer`, `cousin`, `dance`, `day`, `deaf`,
93
- `doctor`, `dog`, `draw`, `drink`, `eat`, `english`, `family`, `father`, `fine`, `finish`,
94
- `fish`, `forget`, `friend`, `girl`
95
-
96
- ## 🚀 使用說明
97
- 1. 上傳影像或使用攝像頭拍攝手語動作
98
- 2. 系統會自動辨識並顯示結果
99
- 3. 點擊「清除預測」重新開始
100
- """, elem_classes=["header"])
101
-
102
- with gr.Row():
103
- with gr.Column(scale=2):
104
- # 影像輸入 - 使用 Image 組件解決 Gradio 4.44.x 兼容性問題
105
- image_input = gr.Image(
106
- label="📹 影像輸入 (攝像頭/上傳)",
107
- sources=["webcam", "upload"],
108
- height=400
109
- )
110
-
111
- # 處理按鈕
112
- process_btn = gr.Button(
113
- "🔍 分析手語動作",
114
- variant="primary",
115
- size="lg"
116
- )
117
-
118
- # 清除按鈕
119
- clear_btn = gr.Button(
120
- "🗑️ 清除預測序列",
121
- variant="secondary",
122
- size="lg"
123
- )
124
-
125
- with gr.Column(scale=1):
126
- # 處理後的影像輸出
127
- output_image = gr.Image(
128
- label="🎯 辨識結果影像",
129
- height=400
130
- )
131
-
132
- # 預測輸出
133
- prediction_output = gr.Textbox(
134
- label="📊 辨識結果",
135
- lines=8,
136
- value="等待影像輸入...",
137
- elem_classes=["prediction-box"]
138
- )
139
-
140
- # 系統資訊
141
- gr.Markdown("""
142
- ## 📊 系統資訊
143
- - **模型準確率**: 94.25%
144
- - **F1分數**: 94.24%
145
- - **處理速度**: 30 FPS
146
- - **特徵提取**: MediaPipe + 光流
147
- - **模型架構**: BiLSTM + 注意力機制
148
- - **背景分割**: MediaPipe Segmentation
149
-
150
- ## 🔧 技術特色
151
- - MediaPipe Holistic 關鍵點檢測
152
- - 光流動作特徵捕捉
153
- - 人體分割背景去除
154
- - 深度學習時序建模
155
- """)
156
-
157
- # 事件處理器
158
- process_btn.click(
159
- fn=process_video_frame,
160
- inputs=[image_input],
161
- outputs=[output_image, prediction_output]
162
- )
163
-
164
- clear_btn.click(
165
- fn=clear_predictions,
166
- outputs=[prediction_output]
167
- )
168
-
169
- gr.Markdown("""
170
- ---
171
- ### 📈 關於此系統
172
-
173
- SignView2.0 使用最先進的深度學習技術進行手語辨識:
174
-
175
- - **特徵提取**: 使用 MediaPipe 提取手部、身體關鍵點 + 光流特徵
176
- - **背景處理**: MediaPipe Segmentation 自動去除背景干擾
177
- - **時序建模**: 雙向LSTM + GRU + 多頭注意力機制
178
- - **訓練資料**: 2380個手語視頻,經過數據增強和正規化
179
-
180
- **開發者**: XiaoBai1221 | **平台**: Hugging Face Spaces
181
- """)
182
-
183
- # 啟動應用程式 - Hugging Face Spaces 會自動檢測 demo 變數
184
- print("🎉 SignView2.0 手語辨識系統已啟動!")
185
-
186
- # 直接啟動,不使用條件判斷
187
- demo.launch(share=True)