Model compression method and device, target detection equipment and storage medium

Abstract

The invention provides a model compression method and device, target detection equipment and a storage medium. The compression method comprises the following steps: acquiring a sample image, and constructing a standard detection model; setting a regular term of a scaling coefficient of a batch standardization layer in the standard detection model; performing sparse training on the standard detection model according to the regular term; pruning the standard detection model according to the scaling coefficient; and performing fine adjustment processing on the pruned standard detection model to obtain a target detection model. According to the compression method, the detection precision of the target detection model can be guaranteed, the detection speed of the target detection model is greatly increased, the input-output ratio of actual engineering is greatly increased, the compression method can be directly operated on a mature framework, support of a special algorithm library is not needed, and the universality is improved.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a compression method of a target detection model, a compression device of a target detection model, a target detection device and a computer-readable storage medium. Background technique [0002] The deep convolutional network mainly uses end-to-end automatic learning to capture the characteristics of objects through strategies such as weight sharing and local connection, so that the network has stronger analytical capabilities and greatly improves the accuracy of target detection compared with traditional algorithms. Among them, the most representative one-stage (single-stage) target detection model YOLO (You Only Look One) iterates three versions, and YOLOv3 achieves a harmonious unity of speed and accuracy. However, in order to maximize the input-output ratio of the engineering project, a GPU (image processor) server is required to support real-time detection of ...

AI Technical Summary

Problems solved by technology

However, in order to maximize the input-output ratio of the project, a GPU (image processor) server is required to support real-time detection of targets by multiple cameras, and the existing YOLO v3 detection speed (about 16.5ms) is difficult to support real-time detection of multiple cameras

[0003] At present, the following methods are generally used to reduce the scale of the YOLO v3 model and expand its practicability: 1. The low-rank decomposition method can compress the model size by 3 times, but the operation speed has not been significantly improved, and the effect on the convolutional layer is not good; 2. Weight Quantization (Weight Quantization) method, although grouping and shared weight methods can be used to compress the number of parameters required for storage, but there is no compression on the Run-time memory (running memory); 3. Binarization weight 4. Weight pruning / sparsifying (Weight Pruning / Sparsifying) method, based on kernel sparsification for model compression, adding regular term induction to the update of weights, so that most of the weights are It is 0, but after pruning, general-purpose runtime libraries (such as conv2d) cannot be called directly, and a new computing interface needs to be designed separately, which requires dedicated hardware or code library support, and poor versatility

Images