Advanced HOG research: Multi-scale partitioning HOG algorithm

The Histogram of Oriented Gradients algorithm is a widely adopted technique in visual feature extraction, especially in object detection tasks. Traditionally, HOG descriptors rely on fixed-size cells and blocks to capture local gradient distributions. Although this technique has been proven effective in pedestrian detection, its single-scale structure limits its ability to represent features in changing spatial patterns. To address this limitation, this study introduces an improved feature extraction method that integrates multi-scale partitioning. Inspired by feature pyramid representation, the method captures image characteristics in different receptive fields by using multiple combinations of cell and block sizes. These features are aggregated to form amore discriminative and robust representation. Comprehensive experiments on multiple public datasets demonstrate that the proposed multi-scale HOGframework consistently outperforms standard HOGmethods, providing remarkable improvements in both classification and detection accuracy. The results validate the effectiveness and adaptability of the method in diverse visual recognition scenarios.