A Hybrid CNN-Multiclass SVM Model for Household Object Recognition: A System for Domestic Robotic Vision
Smita Gour, Pushpa B. Patil · International Journal of Computing · 2025
Computer vision is a broad area covering many aspects of detection and recognition of objects. Household object recognition is a challenging problem, mainly due to lighting conditions, differences in shape, size of objects, position, occlusion, clutter of the objects and background. A novel deep learning and multiclass Support Vector Machine (SVM) based mechanism for household object recognition, is presented here. A comprehensive dataset of household objects, Washington RGB-D, is used. The Convolutional Neural Network (CNN) is used for feature extraction. It involves building and training CNN model using RGB-D dataset. Once CNN model is trained the features are automatically extracted from fully connected layer. The normalized feature vectors are used as input to train the SVM classifier. Empirical analysis of the proposed system's performance is carried out using RGB-D and real-time datasets. This hybrid model classifies the real-time instances with an accuracy of 89%. It exhibited an accuracy of 90.2% with RGB-D standard dataset, taken alone. The robustness and accuracy of the proposed approach pave the way for improved interaction between robots and humans in interior environments.