Research And Application Of Model Quantitative Methods In Computer Vision

With the rapid development of deep learning,the model size has also increased.Due to the need for real-time deployment of computer vision algorithms,mobile terminal reasoning needs to have the ability of fast calculation and the need to save energy consumption,and research and application of model compression have also emerged.Model quantization is an effective strategy for deep neural network model compression.Bit operations have great potential in improving the speed of inference.However,at present,when the quantization method widely used in the open source inference framework is quantized to a low bit,the accuracy of the quantized model is still not small compared to the floating-point model.In order to make up as much as possible for the accuracy loss caused by quantized deep neural networks,this paper has carried out research on quantization algorithms on several computer vision algorithms.The main direction of model quantization algorithm experiments in this paper is the quantization scheme based on fine-tuning.In the selection of computer vision algorithms,the light classification network SQUEEZENET v1.1 and object detection network YOLO v2-SQUEEZENET v1.1,which are rarely applied by domestic and foreign quantization algorithms,are selected.When quantizing deep neural networks to lower bits,incremental quantization is used as the training tricks,specifically the combination of hierarchical quantization and two-step quantization to try to recover the accuracy loss caused by quantization.The structure of the two-dimensional Fourier transform method found on the correlation filter tracking algorithm CCSK can also be applied with the quantization method idea.After proving that the complex matrix operation can also use the fixed-point idea of model quantization,the two-dimensional Fourier transform method is quantified using the quantization algorithms RISTRETTO and IAO.In the experiments in this paper,we investigate and refine the current fine-tuning-based quantization methods at home and abroad,and try to make the quantization process simple,but still maintain good accuracy.From the final experiments,the light-weight network SQUEEZENET v1.1 and the object detection network YOLOv2-SQUEEZENET v1.1 both loss less precision in quantization.Preliminary fixed-point testing and application have also been completed on the single-target tracking algorithm CCSK.