Research on Lightweight Object Detection Algorithm for Hand Detection

In the task of hand-object detection, tradition nal object detection models often face challenges such as large sizes and high parameter counts, making them unsuitable for devices with limited storage and computational resources. To address these issues, this paper proposes an improved YOLOv3 algorithm. First, structured channel pruning is applied to the trained model based on the λ coefficient, pruning channels with λ values close to zero, which have minimal impact on model performance. This significantly reduces both the model size and the number of parameters. Then, the pruned model is further compressed using the TensorRT tool for half-precision quantization, converting the FP32 model to FP16, further reducing the model size. Experimental results on the Oxford Hand public dataset show that the compressed model only requires 19.4MB of storage, representing a 91.8% reduction compared to the original model's 235MB. The number of parameters is reduced by 83.6%, down to just 10.08M. Moreover, the compressed model maintains a high level of accuracy with a mean average precision (mAP) of 62.67%, showing almost no reduction in accuracy compared to the uncompressed model.