In class-incremental semantic segmentation (CISS), models suffer from catastrophic forgetting and background shift. These challenges cause severe semantic confusion among old classes, new classes, and the background. They also lead to feature representation bias. Existing intermediate-layer feature distillation methods typically enforce consistency only on old-class and background regions through regularization. They overlook the interactive relationships among old-class, new-class, and background features. As a result, discriminative semantic relationships are insufficiently preserved. Insufficient classifier optimization further leads to classifier collapse. Traditional pseudo-labeling techniques commonly rely on entropy minimization or high-confidence thresholding. These methods are sensitive to dataset class distributions. They are also prone to overfitting. Such limitations are fundamental causes of class confusion and feature bias. To address these issues, we propose Learning Discriminative Representations to Mitigate Class Confusion in Class-Incremental Semantic Segmentation (LDRMCC). The proposed framework consists of three components. Neighborhood Relationship Knowledge Distillation (NRKD) explicitly models spatial-semantic relationships. It constructs affinity graphs from old and new models. It distills relationships among old-class, new-class, and background features. Semantic Prototype Cross-Entropy Loss (SPCE) improves classifier optimization. It computes inter-prototype similarities. It introduces a dynamic threshold weighting mechanism. This design emphasizes hard-to-separate classes and alleviates class confusion. Normalized Variance Pseudo-Labeling (NVPL) enhances pseudo-label quality. It integrates top-1 and top-2 prediction confidences. It introduces an ignore index. It employs a normalized pseudo-variance metric in the range [0,1]. This strategy adapts to different datasets and generates reliable pseudo-labels. Extensive experiments on the Pascal VOC 2012 and ADE20K datasets demonstrate the effectiveness of LDRMCC. The proposed method achieves state-of-the-art performance under various incremental settings.

@article{liu2025learning,
  title={Learning Discriminative Representations to Mitigate Class Confusion in Class-Incremental Semantic Segmentation},
  author={Liu, Weixiang and Yang, Pan and He, Liangji and Zeng, Deyu and Iqbal, Saeed and Wu, Zongze and Zhong, Xiaopin},
  journal={Pattern Recognition},
  note={In preparation},
  year={2026}
}

An on-device keyword spotting system for detecting pet vocalizations, built on the Seeed XIAO ESP32S3 Sense board. The system leverages TinyML techniques to perform real-time audio inference on a resource-constrained microcontroller, recognizing specific pet sounds (e.g., barking, meowing) without cloud connectivity. The project explores KWS (Keyword Spotting) model training, on-device deployment, and edge computing for smart home applications — enabling responsive, privacy-preserving pet monitoring at the edge.

With the rapid development of Internet of Things applications, home security systems are increasingly required to provide low latency, reliable warning, low power consumption, and better privacy protection. Traditional cloud-centered security architectures usually upload raw sensor data to remote servers, which may introduce network delay, bandwidth pressure, and potential privacy risks. To address these problems, this paper proposes and implements a lightweight edge-computing-based smart home security system. The system uses an ESP32-S3 edge node to collect sound signals and perform local inference. Mel Frequency Energy (MFE) features are extracted from one-second audio segments and fed into a lightweight one-dimensional convolutional neural network. Instead of transmitting raw audio, the edge node publishes only classification results and confidence values to a Message Queuing Telemetry Transport (MQTT) broker. Home Assistant then subscribes to the corresponding topics and executes hierarchical automation rules, including local alarm activation, smart lock control, and mobile notification. Experimental results show that the proposed MFE + Conv1D model achieves 97.1% validation accuracy. In a simulated home environment, the average recognition accuracy reaches 92.5%, the single-sample inference latency is less than 20 ms, and the emergency alarm accuracy reaches 96% under compound dog-bark and glass-breaking scenarios. The results demonstrate that the proposed system can provide a low-cost, low-latency, and privacy-aware solution for practical smart home security.