Intelligent Diagnosis of Urban Underground Drainage Network: From Detection to Evaluation

Table 1

Advantages and disadvantages of typical convolutional neural structure.

Category	Network structure	Year of appearance	Advantages	Disadvantages	Application
Image classification	AlexNet	2012	Outstanding performance in the ImageNet image classification competition	A large number of parameters, easy to over-fitting	[3, 6]
	VGGNet	2014	The model has a clear hierarchy and excellent performance	The parameter quantity is large, and the model size is large	[59]
	GoogLeNet	2014	The number of parameters is relatively small, and the accuracy is high	The complex network structure is complex	[60]
	ResNet	2015	Can solve the problem of deep network gradient disappearance and explosion	The parameter quantity and calculation amount are large	Widely applied
Target detection	R-CNN	2014	Use convolution to extract features for target detection	A large amount of calculation, long time-consuming	—
	SPP-Net	2014	Can process different image sizes, and the recognition speed is fast	Many features need to be stored and long training time	[61]
	Fast R-CNN	2015	Improve the speed and performance of R-CNN	Cannot get rid of dependence on the constituency	—
	Faster R-CNN	2015	Further improves the speed, has advantages in high precision detection	Two-stage network, unable to detect in real time	[62]
	YOLO	2015	Can ensure accuracy while real-time detection	Insufficient ability to identify small targets	[7]
	SSD	2016	Fast speed, high precision, and excellent comprehensive performance	Many parameters need to be manually set	[63]
Pixel-level segmentation	FCN	2015	Segmentation of pixel points replaces traditional classification network	Segmentation is not refined enough and requires a large storage capacity	[64]
	U-Net	2015	Partially solve the problem of insufficient memory and insufficient training data	Difficult to deal with the segmentation tasks of complex scenes and small targets, and prone to overfitting	[65]
	SegNet	2015	Upsampling sharpens edges and reduces storage space	Unable to handle spatial relationships well	[66]
	Mask R-CNN	2017	Second-stage pixel-level localization to achieve instance segmentation	Detecting and segmenting takes a long time	[23]
	DeepLab3+	2018	Proposing dilated convolutions and ASPP modules for high accuracy	Training and processing speeds are slow	[67]
	SOLO	2020	Efficiently recognize multiscale objects and perform pixel-level segmentation	Weak ability to deal with complex backgrounds and not precise enough in detecting small objects	[68]

Note. “—” indicates that this algorithm is no longer considered a key module or innovative point in recent research on pipeline detection and is either not used or only used as a feature extractor for comparison with other algorithms in practical applications.