Essence: Set different grayscale thresholds and classify the image grayscale histogram (the same grayscale range belongs to the same category and has a certain similarity)

process:

Classification:

Analyzed:

Theoretical basis: The gray value of the boundary pixel is greatly different from the gray value of the adjacent pixel.

Process: Connect points (edge ​​points) with large differences in gray value from adjacent pixels to form a boundary outline

Classification:

Summary: In practical applications, the parallel edge detection method is simple and fast, has relatively good performance, and is the most commonly used method.

Algorithm principle: Segment according to image spatial information, classify pixels and form regions through the similarity features of pixels

Classification

Algorithm principle: gather pixels with similar characteristics into the same area, iterate the clustering results repeatedly until convergence, and finally gather all pixels into several different categories to complete image area division == image segmentation

Example analysis of typical algorithms

Algorithm idea: convert the segmentation problem into graph partitioning, and complete the segmentation by optimizing the solution of the objective function

Examples of classic algorithms

mathematical morphology theory

genetic algorithm

Wavelet transform

active contour model

fuzzy theory

rough set theory

... ...

After 8 layers of convolution processing, the feature map is upsampled to implement a deconvolution operation, classified through the SoftMax layer, and finally the segmentation result is output - multiple convolution operations. The size of the feature map is much smaller than the original input image, and many underlying features are lost. Image information, directly classified, affects segmentation accuracy

The upsampling process adopts Skip strategy

Combine deep data with shallow information, and then restore the output of the original image to obtain more accurate segmentation results.

According to different pooling layers, it is divided into

FCN-8s: Integrate more layers of feature information, segment to obtain clearer contour information, and the segmentation effect is relatively good.

It can classify images at the pixel level and effectively solve the problem of image semantic segmentation.

Images of any size can be input

The first end-to-end segmentation network model

insufficient

Integrate contextual information, make full use of prior knowledge of global features, analyze different scenes, and achieve semantic segmentation of scene targets.

1. Given an input image

2.CNN: Obtain the convolutional layer feature map

3. Pyramid pooling module: collect features of different sub-intervals

4. Upsampling

5. Concatenate and fuse the features of each sub-region

6. Form feature representations containing local and global context information

7. Convolution and SoftMax classification of feature representations

8. Prediction results for each pixel

For scene parsing and semantic segmentation tasks - able to extract appropriate global features

Use the pyramid pooling module to fuse local and global information together

Propose an optimization strategy for moderate supervision loss

Disadvantages: The handling of occlusion between targets is not ideal.

Explicitly control the resolution of the response when computing characteristic responses

Expand the receptive field of the convolution kernel

Integrate more feature information without increasing the amount of parameters and calculations

The earliest DeepLab model

DeepLab-v2 model

DeepLab-v3 model

DeepLad-v3 model-encoding and decoding structure

Algorithm framework

Training phase: using multi-task loss constraint L

On the basis of semantic segmentation, instance segmentation is realized - accurate detection and positioning of foreground targets, distinguishing different individuals of similar targets.

Higher segmentation accuracy

Models are more flexible

Can be used for a variety of computer vision tasks

PASCAL VOC

MicrosoftCOCO

Cityscapes

Semantic segmentation: The average intersection and union ratio mIoU represents the ratio of the intersection and union of two sets. In semantic segmentation, it refers to the set of true values ​​and predicted values.

Instance segmentation: Pixel accuracy PA, which represents the proportion of correctly classified pixels to the total pixels

Image segmentation is increasingly used in computer vision tasks

Accuracy and speed have been significantly improved

Lack of segmentation data sets and heavy annotation work

Small-sized target segmentation is not accurate enough

The segmentation algorithm is computationally complex

Unable to achieve real-time interactive segmentation, hindering the implementation, application and promotion of segmentation technology