首先我们将采集的视频文件解码抽取关键帧，每隔4帧保存一次，然后对图像进行中心裁剪和预处理，代码如下：

def load_video(file_name): cap = cv2.VideoCapture(file_name) # 每隔多少帧抽取一次 frame_interval = 4 frames = [] count = 0 while True: ret, frame = cap.read() if not ret: break # 每隔frame_interval帧保存一次 if count % frame_interval == 0: # 中心裁剪 frame = crop_center_square(frame) # 缩放 frame = cv2.resize(frame, (IMG_SIZE, IMG_SIZE)) # BGR -> RGB [0,1,2] -> [2,1,0] frame = frame[:, :, [2, 1, 0]] frames.append(frame) count += 1 return np.array(frames)

然后我们创建图像特征提取器，使用预训练模型InceptionResNetV2提取图像特征，代码如下：

def get_feature_extractor(): feature_extractor = keras.applications.inception_resnet_v2.InceptionResNetV2( weights = 'imagenet', include_top = False, pooling = 'avg', input_shape = (IMG_SIZE, IMG_SIZE, 3) ) preprocess_input = keras.applications.inception_resnet_v2.preprocess_input inputs = keras.Input((IMG_SIZE, IMG_SIZE, 3)) preprocessed = preprocess_input(inputs) outputs = feature_extractor(preprocessed) model = keras.Model(inputs, outputs, name = 'feature_extractor') return model

接着提取视频特征向量，如果视频不足40帧就创建全0数组进行补白：

def load_data(videos, labels): video_features = [] for video in tqdm(videos): frames = load_video(video) counts = len(frames) # 如果帧数小于MAX_SEQUENCE_LENGTH if counts < MAX_SEQUENCE_LENGTH: # 补白 diff = MAX_SEQUENCE_LENGTH - counts # 创建全0的numpy数组 padding = np.zeros((diff, IMG_SIZE, IMG_SIZE, 3)) # 数组拼接 frames = np.concatenate((frames, padding)) # 获取前MAX_SEQUENCE_LENGTH帧画面 frames = frames[:MAX_SEQUENCE_LENGTH, :] # 批量提取特征 video_feature = feature_extractor.predict(frames) video_features.append(video_feature) return np.array(video_features), np.array(labels)

最后创建VIT Model，代码如下：

# 位置编码 class PositionalEmbedding(layers.Layer): def __init__(self, seq_length, output_dim): super().__init__() # 构造从0~MAX_SEQUENCE_LENGTH的列表 self.positions = tf.range(0, limit=MAX_SEQUENCE_LENGTH) self.positional_embedding = layers.Embedding(input_dim=seq_length, output_dim=output_dim) def call(self,x): # 位置编码 positions_embedding = self.positional_embedding(self.positions) # 输入相加 return x + positions_embedding # 编码器 class TransformerEncoder(layers.Layer): def __init__(self, num_heads, embed_dim): super().__init__() self.p_embedding = PositionalEmbedding(MAX_SEQUENCE_LENGTH, NUM_FEATURES) self.attention = layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim, dropout=0.1) self.layernorm = layers.LayerNormalization() def call(self,x): # positional embedding positional_embedding = self.p_embedding(x) # self attention attention_out = self.attention( query = positional_embedding, value = positional_embedding, key = positional_embedding, attention_mask = None ) # layer norm with residual connection output = self.layernorm(positional_embedding + attention_out) return output def video_cls_model(class_vocab): # 类别数量 classes_num = len(class_vocab) # 定义模型 model = keras.Sequential([ layers.InputLayer(input_shape=(MAX_SEQUENCE_LENGTH, NUM_FEATURES)), TransformerEncoder(2, NUM_FEATURES), layers.GlobalMaxPooling1D(), layers.Dropout(0.1), layers.Dense(classes_num, activation="softmax") ]) # 编译模型 model.compile(optimizer = keras.optimizers.Adam(1e-5), loss = keras.losses.SparseCategoricalCrossentropy(from_logits=False), metrics = ['accuracy'] ) return model

模型训练

完整体验可以点击Run in ModelArts一键运行我发布的Notebook：

最终模型在整个数据集上的准确率达到87%，即在小数据集上训练取得了较为不错的结果。

视频推理

首先加载VIT Model，获取视频类别索引标签：

import random # 加载模型 model = tf.keras.models.load_model('saved_model') # 类别标签 label_to_name = {0:'无效手势', 1:'上滑', 2:'下滑', 3:'左滑', 4:'右滑', 5:'打开', 6:'关闭', 7:'放大', 8:'缩小'}

然后使用图像特征提取器InceptionResNetV2提取视频特征：

# 获取视频特征 def getVideoFeat(frames): frames_count = len(frames) # 如果帧数小于MAX_SEQUENCE_LENGTH if frames_count < MAX_SEQUENCE_LENGTH: # 补白 diff = MAX_SEQUENCE_LENGTH - frames_count # 创建全0的numpy数组 padding = np.zeros((diff, IMG_SIZE, IMG_SIZE, 3)) # 数组拼接 frames = np.concatenate((frames, padding)) # 取前MAX_SEQ_LENGTH帧 frames = frames[:MAX_SEQUENCE_LENGTH,:] # 计算视频特征 N, 1536 video_feat = feature_extractor.predict(frames) return video_feat

最后将视频序列的特征向量输入Transformer Encoder进行预测：

# 视频预测 def testVideo(): test_file = random.sample(videos, 1)[0] label = test_file.split('_')[-2] print('文件名:{}'.format(test_file) ) print('真实类别:{}'.format(label_to_name.get(int(label))) ) # 读取视频每一帧 frames = load_video(test_file) # 挑选前帧MAX_SEQUENCE_LENGTH显示 frames = frames[:MAX_SEQUENCE_LENGTH].astype(np.uint8) # 保存为GIF imageio.mimsave('animation.gif', frames, duration=10) # 获取特征 feat = getVideoFeat(frames) # 模型推理 prob = model.predict(tf.expand_dims(feat, axis=0))[0] print('预测类别：') for i in np.argsort(prob)[::-1][:5]: print('{}: {}%'.format(label_to_name[i], round(prob[i]*100, 2))) return display(Image(open('animation.gif', 'rb').read()))

模型预测结果：

文件名:hand_gesture/woman_014_0_7.mp4 真实类别:无效手势 预测类别： 无效手势: 99.82% 下滑: 0.12% 关闭: 0.04% 左滑: 0.01% 打开: 0.01%

 

点击关注，第一时间了解华为云新鲜技术~