OpenCV4 部署DeepLabv3+模型

引言 ·

前面说了OpenCV DNN不光支持图像分类与对象检测模型。此外还支持各种自定义的模型，deeplabv3模型是图像语义分割常用模型之一，本文我们演示OpenCV DNN如何调用Deeplabv3模型实现图像语义分割，支持的backbone网络分别为MobileNet与Inception。预训练模型下载地址如下：

https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md

预训练的模型下载之后可以看到pb文件，ckpt文件，其中pb文件可以直接调用。

调用MobileNet的deeplabv3

下载MobileNet版本的deeplabv3模型，把mobilenetv2 ckpt转pb，脚本如下：

python deeplab/export_model.py 
--logtostderr 
--checkpoint_path="/home/lw/data/cityscapes/train/model.ckpt-2000" 
--export_path="/home/lw/data/pb/frozen_inference_graph.pb" 
--model_variant="mobilenet_v2" 
#--atrous_rates=6 
#--atrous_rates=12 
#--atrous_rates=18 
#--output_stride=16 
--decoder_output_stride=4 
--num_classes=6 
--crop_size=513 
--crop_size=513 
--inference_scales=1.0

接下来使用opencv加载mobilenetv2转换好的pb模型会报下面的错误：

使用mobilenetv2的解决办法:

import tensorflow as tf
from tensorflow.tools.graph_transforms import TransformGraph
from tensorflow.python.tools import optimize_for_inference_lib


graph = 'frozen_inference_graph.pb'
with tf.gfile.FastGFile(graph, 'rb') as f:
     graph_def = tf.GraphDef()
     graph_def.ParseFromString(f.read())
     tf.summary.FileWriter('logs', graph_def)
     inp_node = 'MobilenetV2/MobilenetV2/input'
     out_node = 'logits/semantic/BiasAdd'
     graph_def = optimize_for_inference_lib.optimize_for_inference(graph_def, [inp_node], [out_node],
                                                                   tf.float32.as_datatype_enum)
     graph_def = TransformGraph(graph_def, [inp_node], [out_node], ["sort_by_execution_order"])


with tf.gfile.FastGFile('frozen_inference_graph_opt.pb', 'wb') as f:
     f.write(graph_def.SerializeToString())

调用Inception的deeplabv3 接下来使用opencv加载xception转换好的pb模型

使用xception的解决办法

import tensorflow as tf
from tensorflow.tools.graph_transforms import TransformGraph
from tensorflow.python.tools import optimize_for_inference_lib
graph = 'frozen_inference_graph.pb'
with tf.gfile.FastGFile(graph, 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    tf.summary.FileWriter('logs', graph_def)
    # inp_node = 'sub_2'  # 起始地节点
    # out_node = 'logits/semantic/BiasAdd' # 结束的节点
    graph_def = optimize_for_inference_lib.optimize_for_inference(graph_def, [inp_node], [out_node],
                                                                  tf.float32.as_datatype_enum)
    graph_def = TransformGraph(graph_def, [inp_node], [out_node], ["sort_by_execution_order"])

with tf.gfile.FastGFile('frozen_inference_graph_opt.pb', 'wb') as f:
    f.write(graph_def.SerializeToString())

使用opencv进行推理（我用的是ADE20K预训练模型进行测试的）

import cv2
import numpy as np
np.random.seed(0)
color = np.random.randint(0, 255, size=[150, 3])
print(color)
# Load names of classes
#classes = None


#with open("labels.names", 'rt') as f:
#    classes = f.read().rstrip('n').split('n')


#legend = None
#def showLegend(classes):
#    global legend
#    if not classes is None and legend is None:
#        blockHeight = 30
#        print(len(classes), len(colors))
#        assert(len(classes) == len(colors))
#        legend = np.zeros((blockHeight * len(colors), 200, 3), np.uint8)
#        for i in range(len(classes)):
#            block = legend[i * blockHeight:(i + 1) * blockHeight]
#            block[:, :] = colors[i]
#            cv2.putText(block, classes[i], (0, blockHeight//2), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255))
#        cv2.namedWindow('Legend', cv2.WINDOW_NORMAL)
#        cv2.imshow('Legend', legend)
#        cv2.waitKey()


# 读取图片
frame = cv2.imread("1.jpg")
frameHeight = frame.shape[0]
frameWidth = frame.shape[1]


# 加载模型
net = cv2.dnn.readNet("frozen_inference_graph_opt.pb")
blob = cv2.dnn.blobFromImage(frame, 0.007843, (513, 513), (127.5, 127.5, 127.5), swapRB=True)
net.setInput(blob)
score = net.forward()
numClasses = score.shape[1]
height = score.shape[2]
width = score.shape[3]

classIds = np.argmax(score[0], axis=0)  # 在列上求最大的值的索引
segm = np.stack([color[idx] for idx in classIds.flatten()])
segm = segm.reshape(height, width, 3)

segm = cv2.resize(segm, (frameWidth, frameHeight), interpolation=cv2.INTER_NEAREST)
frame = (0.3*frame + 0.8*segm).astype(np.uint8)

#showLegend(classes)

cv2.imshow("img", frame)
cv2.waitKey()