分割對于圖像解釋任務至關重要,那就不要落後于流行趨勢,讓我們來實施它,我們很快就會成為專業人士!
什麼是語義分割?
它描述了将圖像的每個像素與類别标簽(例如花、人、道路、天空、海洋或汽車)相關聯的過程,即我們要輸入圖像,然後為該圖像中的每個像素輸出一個類别決策。例如下面這個輸入圖像,這是一隻坐在床上的狗:
是以,在輸出中,我們希望為每個像素定義一組類别,即狗、床、後面的桌子和櫥櫃。在語義分割之後,圖像看起來像這樣:
關于語義分割的一件有趣的事情是它不區分執行個體,即如果此圖像中有兩隻狗,它們将僅被描述為一個标簽,即 dog ,而不是 dog1 和 dog2。
語義分割一般用于:
- 自動駕駛
- 工業檢驗
- 衛星圖像中值得注意的區域分類
- 醫學影像監查
語義分割實作:
- 第一種方法是滑動視窗,我們将輸入圖像分解成許多小的局部圖像,但是這種方法在計算上會很昂貴。是以,我們在實踐中并沒有真正使用這個方法。
- 另一種方法是完全卷積網絡,其中網絡有一整堆卷積層,沒有完全連接配接的層,進而保留了輸入的空間大小,這在計算上也是極其昂貴的。
-
第三個也是最好的一個方法,那就是對圖像進行上采樣和下采樣。是以,我們不需要對圖像的完整空間分辨率進行所有卷積,我們可能會在原始分辨率下周遊少量卷積層,然後對該特征圖進行下采樣,然後對其進行上采樣。
在這裡,我們隻想在網絡的後半部分提高我們預測的空間分辨率,以便我們的輸出圖像現在可以與我們的輸入圖像具有相同的次元。它的計算效率要高得多,因為我們可以使網絡非常深,并以更便宜的空間分辨率運作。
讓我們在代碼中實作這一點:
-
導入處理所需的必要庫,即
Pytorch 的重要功能,例如資料加載器、變量、轉換和優化器相關函數。
導入 VOC12 和 cityscapes 的資料集類,從 transform.py 檔案導入 Relabel、ToLabel 和 Colorize 類,從 iouEval.py 檔案中導入 iouEval 類。
#SSCV IIITH 2K19
import random
import time
import numpy as np
import torch
print(torch.__version__)
import math
from PIL import Image, ImageOps
from torch.optim import SGD, Adam, lr_scheduler
from torch.autograd import Variable
from torch.utils.data import DataLoader
from torchvision.transforms import Resize
from torchvision.transforms import ToTensor, ToPILImage
from dataset import cityscapes
from dataset import idd_lite
import sys
print(sys.executable)
from transform import Relabel, ToLabel, Colorize
import matplotlib
from matplotlib import pyplot as plt
%matplotlib inline
import importlib
from iouEval import iouEval, getColorEntry #importing iouEval class from the iouEval.py file
from shutil import copyfile
複制
- 定義幾個全局參數:
NUM_CHANNELS = 3 #RGB Images
NUM_CLASSES = 8 #IDD Lite has 8 labels or Level1 hierarchy of labels
USE_CUDA = torch.cuda.is_available()
IMAGE_HEIGHT = 160
DATA_ROOT = ‘/tmp/school/6-segmentation/user/1/6-segmentation/idd1_lite’
BATCH_SIZE = 2
NUM_WORKERS = 4
NUM_EPOCHS = 100
ENCODER_ONLY = True
device = torch.device(“cuda” )
#device = ‘cuda’
color_transform = Colorize(NUM_CLASSES)
image_transform = ToPILImage()
IOUTRAIN = False
IOUVAL = True
複制
- 增強,即對圖像和目标執行随機增強的不同功能:
class MyCoTransform(object):
def __init__(self, enc, augment=True, height=160):
self.enc=enc
self.augment = augment
self.height = height
pass
def __call__(self, input, target):
# Resizing data to required size
input = Resize((self.height,320), Image.BILINEAR)(input)
target = Resize((self.height,320), Image.NEAREST)(target)
if(self.augment):
# Random horizontal flip
hflip = random.random()
if (hflip < 0.5):
input = input.transpose(Image.FLIP_LEFT_RIGHT)
target = target.transpose(Image.FLIP_LEFT_RIGHT)
#Random translation 0–2 pixels (fill rest with padding)
transX = random.randint(0, 2)
transY = random.randint(0, 2)
input = ImageOps.expand(input, border=(transX,transY,0,0), fill=0)
target = ImageOps.expand(target, border=(transX,transY,0,0), fill=7) #pad label filling with 7
input = input.crop((0, 0, input.size[0]-transX, input.size[1]-transY))
target = target.crop((0, 0, target.size[0]-transX, target.size[1]-transY))
input = ToTensor()(input)
target = ToLabel()(target)
target = Relabel(255,7)(target)
return input, target
複制
- 加載資料:我們将遵循 pytorch 推薦的語義,并使用資料加載器加載資料。
best_acc = 0
co_transform = MyCoTransform(ENCODER_ONLY, augment=True, height=IMAGE_HEIGHT)
co_transform_val = MyCoTransform(ENCODER_ONLY, augment=False, height=IMAGE_HEIGHT)
#train data
dataset_train = idd_lite(DATA_ROOT, co_transform, ‘train’)
print(len(dataset_train))
#test data
dataset_val = idd_lite(DATA_ROOT, co_transform_val, ‘val’)
print(len(dataset_val))
loader_train = DataLoader(dataset_train, num_workers=NUM_WORKERS, batch_size=BATCH_SIZE, shuffle=True)
loader_val = DataLoader(dataset_val, num_workers=NUM_WORKERS, batch_size=BATCH_SIZE, shuffle=False)
複制
- 既然是分類問題,我們就使用交叉熵損失,但為什麼呢?
答案是負對數,在較小值的時候效果不好,并且在較大值的時候效果也不好。因為我們将損失函數加到所有正确的類别上,實際發生的情況是,每當網絡為正确的類别,配置設定高置信度時,損失就低,但是當網絡為正确的類别時配置設定低置信度,損失就高。
criterion = torch.nn.CrossEntropyLoss()
複制
- 現在讓我們加載模型并優化它!
model_file = importlib.import_module(‘erfnet’)
model = model_file.Net(NUM_CLASSES).to(device)
optimizer = Adam(model.parameters(), 5e-4, (0.9, 0.999), eps=1e-08, weight_decay=1e-4)
start_epoch = 1
複制
- 是以,編碼的最終本質就是訓練!
import os
steps_loss = 50
my_start_time = time.time()
for epoch in range(start_epoch, NUM_EPOCHS+1):
print(“ — — — TRAINING — EPOCH”, epoch, “ — — -”)
epoch_loss = []
time_train = []
doIouTrain = IOUTRAIN
doIouVal = IOUVAL
if (doIouTrain):
iouEvalTrain = iouEval(NUM_CLASSES)
model.train()
for step, (images, labels) in enumerate(loader_train):
start_time = time.time()
inputs = images.to(device)
targets = labels.to(device)
outputs = model(inputs, only_encode=ENCODER_ONLY)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
loss = criterion(outputs, targets[:, 0])
loss.backward()
optimizer.step()
epoch_loss.append(loss.item())
time_train.append(time.time() — start_time)
if (doIouTrain):
#start_time_iou = time.time()
iouEvalTrain.addBatch(outputs.max(1)[1].unsqueeze(1).data, targets.data)
#print (“Time to add confusion matrix: “, time.time() — start_time_iou)
# print statistics
if steps_loss > 0 and step % steps_loss == 0:
average = sum(epoch_loss) / len(epoch_loss)
print(‘loss: {average:0.4} (epoch: {epoch}, step: {step})’, “// Avg time/img: %.4f s” % (sum(time_train) / len(time_train) / BATCH_SIZE))
average_epoch_loss_train = sum(epoch_loss) / len(epoch_loss)
iouTrain = 0
if (doIouTrain):
iouTrain, iou_classes = iouEvalTrain.getIoU()
iouStr = getColorEntry(iouTrain)+’{:0.2f}’.format(iouTrain*100) + ‘\033[0m’
print (“EPOCH IoU on TRAIN set: “, iouStr, “%”)
my_end_time = time.time()
print(my_end_time — my_start_time)
複制
在訓練了 100 個 epoch 之後,我們會看到:
- 驗證:
#Validate on val images after each epoch of training
print(“ — — — VALIDATING — EPOCH”, epoch, “ — — -”)
model.eval()
epoch_loss_val = []
time_val = []
if (doIouVal):
iouEvalVal = iouEval(NUM_CLASSES)
for step, (images, labels) in enumerate(loader_val):
start_time = time.time()
inputs = images.to(device)
targets = labels.to(device)
with torch.no_grad():
outputs = model(inputs, only_encode=ENCODER_ONLY)
#outputs = model(inputs)
loss = criterion(outputs, targets[:, 0])
epoch_loss_val.append(loss.item())
time_val.append(time.time() — start_time)
#Add batch to calculate TP, FP and FN for iou estimation
if (doIouVal):
#start_time_iou = time.time()
iouEvalVal.addBatch(outputs.max(1)[1].unsqueeze(1).data, targets.data)
#print (“Time to add confusion matrix: “, time.time() — start_time_iou)
if steps_loss > 0 and step % steps_loss == 0:
average = sum(epoch_loss_val) / len(epoch_loss_val)
print(‘VAL loss: {average:0.4} (epoch: {epoch}, step: {step})’,
“// Avg time/img: %.4f s” % (sum(time_val) / len(time_val) / BATCH_SIZE))
average_epoch_loss_val = sum(epoch_loss_val) / len(epoch_loss_val)
iouVal = 0
if (doIouVal):
iouVal, iou_classes = iouEvalVal.getIoU()
print(iou_classes)
iouStr = getColorEntry(iouVal)+’{:0.2f}’.format(iouVal*100) + ‘\033[0m’
print (“EPOCH IoU on VAL set: “, iouStr, “%”)
複制
- 可視化輸出:
# Qualitative Analysis
dataiter = iter(loader_val)
images, labels = dataiter.next()
if USE_CUDA:
images = images.to(device)
inputs = images.to(device)
with torch.no_grad():
outputs = model(inputs, only_encode=ENCODER_ONLY)
label = outputs[0].max(0)[1].byte().cpu().data
label_color = Colorize()(label.unsqueeze(0))
label_save = ToPILImage()(label_color)
plt.figure()
plt.imshow(ToPILImage()(images[0].cpu()))
plt.figure()
plt.imshow(label_save)
複制
輸出圖像
很快我們就可以準備好我們的模型了!
随意使用我們新設計的模型,嘗試增加更多的 epoch 并觀察我們的模型表現得更好!
是以,簡而言之,現在我們将能夠輕松地将圖像的每個像素與類标簽相關聯,并可以調整超參數以檢視顯示的更改。本文展示了語義分割的基礎知識,要對執行個體進行分類,我們需要進行執行個體分割,這是語義分割的進階版本。