In the modern era, advanced computer vision technology is tremendously increased. In such a way, image segmentation also becomes an intelligently developing computer vision task. Image segmentation is a critical task in computer vision, a key concept in image processing.
Image segmentation has widespread applications such as scene understanding, medical image analysis, object detection, image retrieval, remote sensing, robotic perception, video surveillance, augmented reality, image compression, and many more.
Image segmentation is divided into two categories: semantic segmentation, instance segmentation, and multitudinous image segmentation algorithms. Due to the immense success of deep learning techniques in a broad range of vision applications for various real-world scenarios, a huge amount of work has been focused on establishing image segmentation approaches utilizing deep learning models.
In computer vision, deep learning technology helps to perform intelligent image segmentation. More recently, Deep Learning models have introduced novel image segmentation models with outstanding performance improvements.
A broad spectrum of deep learning-based image segmentation models is developed, including fully convolutional pixel-labeling networks, encoder-decoder architectures, multi-scale and pyramid-based approaches, recurrent networks, visual attention models, and generative models in adversarial settings. Deep Learning based image segmentation models often accomplish superior accuracy rates on popular benchmarks, resulting in a transformation in the field.
• Fully Convolutional Networks (FCN) - FCN is one of the first deep learning models for semantic image segmentation. Currently, FCNs have been exploited for various segmentation problems, namely brain tumor segmentation, instance-aware semantic segmentation, skin lesion segmentation, and iris segmentation.
• Convolutional Models With Graphical Models - Convolutional Models with Graphical Models are developed for semantic image segmentation by combining convolutional neural networks and Conditional Random Fields. This model also integrates another probabilistic graphical model Markov Random Field.
• Encoder-Decoder-Based Models - Many deep learning-based image segmentation utilizes encoder-decoder-based models based on convolutional encoder-decoder architecture. These models are divided into encoder-decoder models for general segmentation applications and medical image segmentation applications.
• Multi-Scale and Pyramid Network-Based Models - These models adopt multi-scale analysis, an image processing technique that employs several neural network architectures. The leading network structure applied in this model is Feature Pyramid Network. Some applicative models use multi-scale analysis Dynamic Multi-scale Filters Network, Context contrasted network and gated multi-scale aggregation, Adaptive Pyramid Context Networks, Multi-scale context intertwining, and salient object segmentation.
• Regional Convolutional Neural Network (R-CNN) Based Models - R-CNN based models utilized, for instance, image segmentation, and it has a variety of extension models such as Fast R-CNN, Faster R-CNN, Maksed-RCNN. These models are highly used for object detection applications and have reached great heights. Deep Watershed Transform, real-time instance segmentation, and Semantic Instance Segmentation via Deep Metric Learning are some promising research scopes of these models.
• Dilated Convolutional Models and DeepLab Family - Dilated convolutions are a well-known concept in real-time image segmentation. DeepLab family includes various networks such as multi-scale context aggregation, dense up sampling convolution and hybrid dilated convolution, densely connected Atrous Spatial Pyramid Pooling, and the efficient neural network.
• Recurrent Neural Network Based Models - RNNs are beneficial in modeling the short/long term dependencies among pixels conducive to enhancing the estimation of the segmentation map of semantic image segmentation. Gated Recurrent Units (GRUs) and long-short-term-memory (LSTM) networks are architectures applied for RNN-based semantic segmentation.
• Attention-Based Models - Attention mechanisms have been determinedly examined in computer vision applications and explored for semantic image segmentation. Expectation-Maximization Attention (EMANet), Criss-Cross Attention Network (CCNet)-The end-to-end instance segmentation with recurrent attention, a point-wise spatial attention network, and discriminative feature network are some of the networks utilized by attention mechanism for semantic image segmentation.
• Generative Models and Adversarial Training - Generative models are highly applied for computer vision tasks, especially image segmentation. Adversarial networks and generative models are exploited for medical image segmentation, cell image segmentation, segmentation, and generation of the invisible parts of objects.
• CNN Models With Active Contour Models - The integration of FCNs and Active Contour Models (ACMs) has recently emerged and been applied to various medical image segmentation tasks.
• Face detection - Face detection with deep learning-based image segmentation used for biometrics and autofocus features in digital cameras and video sources
• Medical imaging – Recently, deep learning-based medical image analysis has been applied in numerous applications for diagnosing diverse diseases. Medical image segmentation uses deep learning, which helps extract clinically related information from medical images for accurate disease diagnosis.
• Machine vision - In machine vision, deep learning-based image segmentation assists in capturing and processing images, conducive to imparting operational guidance to devices in both industrial and non-industrial applications.
• Video Surveillance - Deep learning-based image segmentation is utilized for video tracking and moving object tracking for security and surveillance, traffic control, human-computer interaction, and video editing.
• Self-driving vehicles -Semantic image segmentation using deep learning enables self-driving cars to recognize safer regions to drive.