Benchmarking Normalization Methods for a CNN Based Object Detection Computer Vision Model
Raidah Binti Yazid, El Filali Sanaa, El Habib Benlahmar · Procedia Computer Science · 2025
Computer vision tasks require precisely chosen components, in this paper we benchmark different data normalization methods, and their impact on convolutional neural networks. The model we focus on for these tests, is one that detects abandoned objects in public spaces that adds safety to public spaces. We evaluate how four normalization methods—Batch Normalization (BN), Layer Normalization (LN), Instance Normalization (IN), and Group Normalization (GN)—influence the model’s classification accuracy and the precision of bounding box localization. We generated a artificial dataset, it replicates real-world security footage, on which the experiments are done. Each normalization method used gives results, that present a useful look at training stability, detection performance as well as the speed of convergence.