Abstract:With remote sensing images constantly being collected rapidly, the class-incremental semantic segmentation (CISS) task has attracted increasing attention. However, the semantic distribution shift problem of the background class in remote sensing images, which is a case of catastrophic forgetting, continues to limit available CISS algorithms. To address this challenge, we present a new memory-inspired semantic segmentation augmentation network (MiSSNet) for class-incremental learning in remote sensing images. The MiSSNet mainly includes two modules: the local semantic distillation (LSD) module and the class-specific regularization (CSR) module. LSD is a distillation structure that employs the local semantic features in retained memory to maintain correlation between pixels throughout the training process of incremental learning. It constructs a series of pixel-level correlation matrices and implicitly adjusts the semantic distribution shift problem of the background class. CSR is a classwise regularization term that utilizes the class-specific portion of the preserved memory to help the model keep repeating the learning of the old categories. It alleviates the background class shift problem by generating countless pixel-level instances of old classes. LSD and CSR work together to tackle the semantic distribution shift problem of background class from semantic information and class information aspects, respectively. In particular, MiSSNet only needs an additional single inference process for memory extraction and storage, and the whole algorithm does not add any new training parameters. Experimental results on three semantic segmentation datasets indicate the advantage of the proposed method.

A Structural Developmental Neural Network with Information Saturation for Continual Unsupervised Learning

Enhancing Efficient Continual Learning with Dynamic Structure Development of Spiking Neural Networks

An enhanced self-organizing incremental neural network for online unsupervised learning.

Efficient Spiking Neural Networks with Sparse Selective Activation for Continual Learning

Deep CovDenseSNN: A Hierarchical Event-Driven Dynamic Framework with Spiking Neurons in Noisy Environment

Design of Incremental Echo State Network Using Leave-One-Out Cross-Validation.

An Incremental Local Distribution Network for Unsupervised Learning.

MiSSNet: Memory-Inspired Semantic Segmentation Augmentation Network for Class-Incremental Learning in Remote Sensing Images

A Multiscale Incremental Learning Network for Remote Sensing Scene Classification

A Self-Organizing Incremental Neural Network Based on Local Distribution Learning

Class Incremental Learning Based on Local Structure Constraints in Feature Space

Autonomous Cognition Development with Lifelong Learning: A Self-Organizing and Reflecting Cognitive Network

Scalable Cognitive Developmental Network: A Strategy for Integrating New Perception Online Using Relation Evolution SOINN

An Incremental Neural Network for Non-stationary Unsupervised Learning

A Self-Structurizing Neural Network for Online Incremental Learning

An Incremental Neural Network for Online Supervised Learning and Topology Learning

ALADE-SNN: Adaptive Logit Alignment in Dynamically Expandable Spiking Neural Networks for Class Incremental Learning

Self-Training and Curriculum Learning Guided Dynamic Refined Network for Remote Sensing Class-Incremental Semantic Segmentation

A Structure-Enriched Neural Network for Network Embedding.

Improved Generalization in Signal Identification With Unsupervised Spiking Neuron Networks for Fiber-Optic Distributed Acoustic Sensor

Class-Incremental Learning for Semantic Segmentation in Aerial Imagery Via Distillation in All Aspects