轻松学Pytorch –车辆类型与颜色识别

OpenCV学堂

发布于 2020-06-29 17:57:42

3.8K1

发布于 2020-06-29 17:57:42

大家好，上一周没有给大家更新这个系列文章，不是我不想更新，而是很多数据需要我自己准备，做好处理，比如这次的车辆属性数据，基于BITVehicle_Dataset公开数据集的基础上，我用程序标注了9000多张车辆属性跟颜色数据集，用于本次训练。本文主要演示了如下一些知识点：

ResNet网络结构的block定义与使用
多分类任务网络设计
OpenVINO Python SDK
多模型推理的先后处理

数据集

前言中交代了，数据来自BITVehicle_Dataset，是一个公开的车辆数据集，从中可以挖掘到很多好玩的数据，它有个文件VehicleInfo.mat，从这个文件中可以获取到车辆的标注信息，每个车辆的ROI区域，车辆类型，我用python读取了这个文件，保存了每个ROI的车辆图像，这样我就得到了车辆属性数据集。其中命名格式如下：

color_type_xxxx.jpg

color表示颜色分类，颜色有7个类别
type 表示车辆类型分类，车型只分了4个类别

1color_labels = ["white", "gray", "yellow", "red", "green", "blue", "black"]
2type_labels = ["car", "bus", "truck", "van"]

同样，通过自定义的Dataset，加载数据集，实现数据集预处理与加载，这部分的内容就不再赘述了，看系列文章的前面相关内容，都有很详细的介绍。

网络模型结构

之前系列文章中给大家演示了卷积神经网络的基本结构跟VGG的stacked卷积的基本结构，这里使用ResNet的Block结构完成了一个简单神经网络，基于该网络实现了对输入车辆图像的颜色与车辆类型的分类，完整的车辆属性识别网络结构如下：

 1VehicleAttributesResNet(
 2  (cnn_layers): Sequential(
 3    (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 4    (1): ReLU()
 5    (2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
 6    (3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
 7    (4): ResidualBlock(
 8      (skip): Sequential(
 9        (0): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
10        (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
11      )
12      (block): Sequential(
13        (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
14        (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
15        (2): ReLU()
16        (3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
17        (4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
18      )
19    )
20    (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
21    (6): ResidualBlock(
22      (skip): Sequential(
23        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
24        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
25      )
26      (block): Sequential(
27        (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
28        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
29        (2): ReLU()
30        (3): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
31        (4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
32      )
33    )
34    (7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
35  )
36  (global_max_pooling): AdaptiveMaxPool2d(output_size=(1, 1))
37  (color_fc_layers): Sequential(
38    (0): Linear(in_features=128, out_features=7, bias=True)
39    (1): Sigmoid()
40  )
41  (type_fc_layers): Sequential(
42    (0): Linear(in_features=128, out_features=4, bias=True)
43  )
44)

其中残差Block卷积的代码实现如下：

 1class ResidualBlock(torch.nn.Module):
 2    def __init__(self, in_channels, out_channels, stride=1):
 3        """
 4        Args:
 5          in_channels (int):  Number of input channels.
 6          out_channels (int): Number of output channels.
 7          stride (int):       Controls the stride.
 8        """
 9        super(ResidualBlock, self).__init__()
10
11        self.skip = torch.nn.Sequential()
12
13        if stride != 1 or in_channels != out_channels:
14            self.skip = torch.nn.Sequential(
15                torch.nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, bias=False),
16                torch.nn.BatchNorm2d(out_channels))
17
18        self.block = torch.nn.Sequential(
19            torch.nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, padding=1, stride=1, bias=False),
20            torch.nn.BatchNorm2d(out_channels),
21            torch.nn.ReLU(),
22            torch.nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, padding=1, stride=1, bias=False),
23            torch.nn.BatchNorm2d(out_channels))
24
25    def forward(self, x):
26        out = self.block(x)
27        identity = self.skip(x)
28        out += identity
29        out = F.relu(out)
30        return out

网络模型的代码实现如下：（上次我把模型代码忘记贴了，这次补上，上一篇代码跟此篇类似、希望大家借此可以解决一系列问题！）

 1class VehicleAttributesResNet(torch.nn.Module):
 2    def __init__(self):
 3        super(VehicleAttributesResNet, self).__init__()
 4        self.cnn_layers = torch.nn.Sequential(
 5            # 卷积层 (64x64x3的图像)
 6            torch.nn.Conv2d(3, 32, 3, padding=1),
 7            torch.nn.ReLU(),
 8            torch.nn.BatchNorm2d(32),
 9            torch.nn.MaxPool2d(2, 2),
10
11            ResidualBlock(32, 64),
12            torch.nn.MaxPool2d(2, 2),
13
14            # 32x32x32
15            ResidualBlock(64, 128),
16            torch.nn.MaxPool2d(2, 2),
17        )
18        # 全局最大池化
19        self.global_max_pooling = torch.nn.AdaptiveMaxPool2d((1, 1))
20        # linear layer (N*9*9*128 ->N*128 )
21
22        self.color_fc_layers = torch.nn.Sequential(
23            torch.nn.Linear(128, 7),
24            torch.nn.Sigmoid()
25        )
26
27        self.type_fc_layers = torch.nn.Sequential(
28            torch.nn.Linear(128, 4),
29        )
30
31    def forward(self, x):
32        # stack convolution layers
33        x = self.cnn_layers(x)
34
35        # 8x8x128
36        B, C, H, W = x.size()
37        out = self.global_max_pooling(x).view(B, -1)
38
39        # 全连接层
40        out_color = self.color_fc_layers(out)
41        out_type = self.type_fc_layers(out)
42        return out_color, out_type

训练与测试

因为两个分支都是分类损失，所以基于交叉熵损失计算、输入的格式是NxCxHxW=Nx16x72x72，我只训练了15个epoch，然后保存模型文件为vehicle_attributes_model.pt。训练的代码如下：

 1# 训练模型的次数
 2num_epochs = 15
 3# optimizer = torch.optim.SGD(model.parameters(), lr=0.001)
 4optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
 5model.train()
 6
 7# 损失函数
 8mse_loss = torch.nn.MSELoss()
 9cross_loss = torch.nn.CrossEntropyLoss()
10index = 0
11for epoch in  range(num_epochs):
12    train_loss = 0.0
13    for i_batch, sample_batched in enumerate(dataloader):
14        images_batch, color_batch, type_batch = \
15            sample_batched['image'], sample_batched['color'], sample_batched['type']
16        if train_on_gpu:
17            images_batch, color_batch, type_batch = images_batch.cuda(), color_batch.cuda(), type_batch.cuda()
18        optimizer.zero_grad()
19
20        # forward pass: compute predicted outputs by passing inputs to the model
21        m_color_out_, m_type_out_ = model(images_batch)
22        color_batch = color_batch.long()
23        type_batch = type_batch.long()
24
25        # calculate the batch loss
26        loss = cross_loss(m_color_out_, color_batch) + cross_loss(m_type_out_, type_batch)
27
28        # backward pass: compute gradient of the loss with respect to model parameters
29        loss.backward()
30
31        # perform a single optimization step (parameter update)
32        optimizer.step()
33
34        # update training loss
35        train_loss += loss.item()
36        if index % 100 == 0:
37            print('step: {} \tTraining Loss: {:.6f} '.format(index, loss.item()))
38        index += 1
39
40        # 计算平均损失
41    train_loss = train_loss / num_train_samples
42
43    # 显示训练集与验证集的损失函数
44    print('Epoch: {} \tTraining Loss: {:.6f} '.format(epoch, train_loss))
45
46# save model
47model.eval()
48torch.save(model, 'vehicle_attributes_model.pt')

然后我使用openvino自带的车辆检测模型，实现车辆检测，在把车辆的ROI区域作为输入，使用训练好的模型，实现了车辆属性识别，最终使用一段视频，验证车辆属性识别的模型，实时运行车辆属性识别结果如下：

实现代码如下：

 1while True:
 2    ret, src = capture.read()
 3    if ret is not True:
 4        break
 5    images = np.ndarray(shape=(n, c, h, w))
 6    images_hw = []
 7    ih, iw = src.shape[:-1]
 8    images_hw.append((ih, iw))
 9    if (ih, iw) != (h, w):
10        image = cv.resize(src, (w, h))
11    image = image.transpose((2, 0, 1))  # Change data layout from HWC to CHW
12    images[0] = image
13    res = exec_net.infer(inputs={input_blob: images})
14
15    # 解析车辆检测输出内容
16    res = res[out_blob]
17    license_score = []
18    license_boxes = []
19    data = res[0][0]
20    index = 0
21    for number, proposal in enumerate(data):
22        if proposal[2] > 0.75:
23            ih, iw = images_hw[0]
24            label = np.int(proposal[1])
25            confidence = proposal[2]
26            xmin = np.int(iw * proposal[3])
27            ymin = np.int(ih * proposal[4])
28            xmax = np.int(iw * proposal[5])
29            ymax = np.int(ih * proposal[6])
30            cv.rectangle(src, (xmin, ymin), (xmax, ymax), (255, 0, 0), 2)
31            if xmin < 0:
32                xmin = 0
33            if ymin < 0:
34                ymin = 0
35            if xmax >= iw:
36                xmax = iw - 1
37            if ymax >= ih:
38                ymax = ih - 1
39
40            # 车辆属性识别
41            vehicle_roi = src[ymin:ymax, xmin:xmax,:]
42            img = cv.resize(vehicle_roi, (72, 72))
43            img = (np.float32(img) / 255.0 - 0.5) / 0.5
44            img = img.transpose((2, 0, 1))
45            x_input = torch.from_numpy(img).view(1, 3, 72, 72)
46            color_, type_ = cnn_model(x_input.cuda())
47            predict_color = torch.max(color_, 1)[1].cpu().detach().numpy()[0]
48            predict_type = torch.max(type_, 1)[1].cpu().detach().numpy()[0]
49            attrs_txt = "color:%s, type:%s"%(color_labels[predict_color], type_labels[predict_type])
50            cv.putText(src, attrs_txt, (xmin, ymin), cv.FONT_HERSHEY_SIMPLEX, 0.75, (0, 0, 255), 2)
51    cv.imshow("Vehicle Attributes Recognition Demo", src)
52    res_key = cv.waitKey(1)
53    if res_key == 27:
54        break

后记：

只说一次的话，其实我每次写文章都会参考一些资料跟连接，但是我这个人比较懒，每次都不会罗列在下面！但是内心是感谢这些国外的分享者的。我真只是参考，绝对不是翻译跟搬运！另外本文虽然说是轻松学Pytorch的系列文章之一，但是涉及到知识点比较多，用到了OpenCV、OpenVINO等开源的视觉框架，所以融会贯通才是王道。轻松学Pytorch系列代码背后的写作一点也不轻松，希望大家点赞支持！

本文参与腾讯云自媒体同步曝光计划，分享自微信公众号。

原始发表：2020-06-25，如有侵权请联系 cloudcommunity@tencent.com 删除

腾讯云开发者社区