A Feature Saliency Based Hybrid Neural Network Model for Object Recognition
Daojie Guo, Yujun Zeng · 2023
Accurate recognition of image targets is a fundamental intelligent perception task and extracting effective features is the prerequisite. However, there are still problems with both hand-crafted features and the ones learned by deep neural networks, such as insufficient generalization ability, limited applicability, and insufficient robustness. This paper first selects representative hand-crafted features and deep convolutional features, analyzing their strengths and weaknesses from the perspective of saliency. Then, inspired by the aforementioned analysis results, a feature saliency extraction process is modeled as the feature coding of the autoencoder based on extreme learning machine (ELM-AE), and further integrated into the object recognition framework composed of a deep convolutional feature extractor and an extreme learning machine classifier. In a consequence, a hybrid neural network model for object recognition based on feature saliency is prposed. Finally, experimental results on the German traffic sign recognition benchmark (GTSRB) show that the proposed model can achieve better performance.