Imagine you can easily photograph and label individual objects, but the real challenge is AI recognizing multiple objects together. A simple solution would be to compile a large dataset with every meaningful object combination. However, for categories with many items, this becomes costly and unfeasible. This patent proposes an alternative: train the model with images of single objects and design it to detect multiple objects during actual use. The process involves passing a multi-object image through a feature-extraction network that creates a three-level image pyramid. Features from different levels are fused, allowing the model to leverage information at various spatial scales for detection. A region-proposal network identifies areas of interest, and the detection network classifies and locates targets. One implementation uses ResNet50 for feature extraction, combined with upsampling and feature fusion. This approach was partly motivated by recognition tasks involving traditional Chinese medicinal materials, where the many categories make multi-object training data hard to produce. Overall, this invention tackles a common issue: training data often doesn’t match real-world conditions. Instead of requiring data for every scenario, the architecture aims to close that gap. The broader significance for practical AI is that creating a “perfect” dataset is usually impossible. A more practical goal is to reorganize available data so that knowledge gained under simpler conditions remains useful in more complex situations.
Training on a single object to recognize multiple
Patent number: CN 114821554 A
Inventor(s): Z. Cai, Z. Jiang, B. Ye, P. Lü, Y. Zhang, J. Liu, T. Lan, and L. Bai
Citation: Z. Cai, Z. Jiang, B. Ye, P. Lü, Y. Zhang, J. Liu, T. Lan, and L. Bai, “图像识别方法、电子设备及存储介质 [Image-recognition method, electronic device and storage medium],” China Patent Application CN 114821554 A, Jul. 29, 2022.
