Optimization method of HarDNet-Lite in embedded platform

Abstract

The invention discloses a method for optimizing HarDNet-Lite on an embedded platform, which is used for solving the problems that the existing target detection network is too complex, the operand is large, the reasoning speed on the embedded platform is low, and the positioning precision is low. The method comprises the following steps: 1) establishing a lightweight HarDNetLite feature extractionnetwork; 2) fusing feature maps of different scales by adopting a weighted FPN structure, so rich underlying detail information and high-level semantic information are fully fused; 3) generating a YOLO detection head, and placing an anchor frame generated through k-means clustering on the feature maps of different sizes to detect targets of different sizes; 4) performing end-to-end model trainingby using a classification and regression loss function; 5) deploying the trained model on an embedded platform, and carrying out target detection. The invention has the beneficial effects that the reasoning speed and the target detection precision of the HarDNet-Lite target detection network on the embedded platform are improved.

Description

technical field [0001] The invention belongs to the technical field of target detection, in particular to an optimization method of HarDNet-Lite on an embedded platform. Background technique [0002] Object detection is an important branch of computer vision, which is widely used in reality, such as video surveillance, industrial detection, face detection and many other fields. Computer vision can reduce the consumption of labor costs, which has a strong practical application significance . [0003] Since 2012, convolutional neural networks have developed rapidly, and some more powerful network structures have been proposed, such as Vgg, GoogleNet, ResNet, ResNext, DenseNet, etc. In order to obtain better performance, the number of layers of the network continues to increase, and the amount of parameters becomes larger and larger. In this way, although the performance of the network has been improved, the problem of efficiency comes with it. Efficiency The problem is mainl...

AI Technical Summary

Problems solved by technology

[0004] (1) Storage problem: With the deepening of the network structure, a large number of convolution kernel weights are stored in the network, so that the storage space occupied by the model itself becomes larger and larger, even up to hundreds of megabytes, while embedded devices The storage space and memory itself are relatively limited, and network models that are too large often cannot be loaded and run;

[0005] (2) Speed ​​problem: In practical applications, the forward inference speed of the model is generally required to be on the order of milliseconds. If the network model is too complex, the amount of calculation will also increase, and the computing power of embedded devices is relatively weak, which will cause the model The speed of forward reasoning becomes very slow, ranging from a few seconds to more than ten seconds, which cannot meet the practical requirements of close to real-time

Images