Mixed-Precision is All You Need for Efficient Document Image Classification

Tushar Shinde, Shivam Bhardwaj; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1285-1293

Efficient document image classification is crucial for resource-constrained applications such as mobile and edge devices. Deploying large-scale neural networks in such environments remains challenging due to computational and memory limitations. To address this we propose an adaptive mixed-precision quantization framework that systematically compresses pre-trained neural networks while preserving their accuracy. We introduce a Layer Quantization Potential metric combining parameter proportion layer entropy and layer standard deviation to rank and quantize layers effectively. We adaptively apply quantization at varying precisions balancing compression and accuracy. Additionally techniques like Huffman encoding are employed for further compression. We evaluate our approach on two benchmark datasets using popular deep neural network architectures such as AlexNet GoogLeNet VGG16 and ResNet50. For the VGG16 model our method achieves compression rates of up to 7.5x (24.3x with Huffman encoding) and 8x (27.9x with Huffman encoding) with minimal degradation in classification performance for Tobacco-3482 and RVL-CDIP-3K datasets respectively. Compared to traditional fixed-precision quantization our adaptive approach provides a superior trade-off between accuracy and model size making it well-suited for real-world deployment.