Method for improving target real-time identification network structure and suitable for embedded equipment

Abstract

The invention discloses a method for improving a target real-time identification network structure and suitable for embedded equipment, and relates to the field of computer vision, and the method specifically comprises the steps: firstly, operating a normal target identification network model, and calculating the proportion of weight parameter quantities of convolution modules of all levels; selecting a convolution module with a proportion of 50%-70%, replacing the convolution module with a depth separable convolution module, and completing primary compression; secondly, introducing the influence factor of each channel into a loss function, and carrying out sparse regularization; setting a cutting proportion threshold, selecting an influence factor lower than the proportion threshold, andcutting off the corresponding network channel to complete a channel pruning and compressing process; then, using an image enhancement module to adjust the image, and inputting the image into the neural network added into the SENet module for training; and finally, performing target identification on the image acquired in real time. According to the method, a structure compression method and a channel pruning method are combined together, so that the original network is compressed to a great extent.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an improved target real-time recognition network structure method suitable for embedded devices. Background technique [0002] With the application of artificial intelligence technology more and more widely, unmanned aerial vehicles (UAV) combined with artificial intelligence will become a hot research field in the future. At present, there have been many applications of transplanting target recognition technology to drones, such as vegetation spraying, disaster search and rescue, and river pollution measurement. Greatly reduce manpower, material and financial resources. In addition, the results of automatic computer identification are often more reliable and safe than human eye identification. [0003] However, in the industry, the process of porting object recognition technology to drones is still full of challenges. For a small embedded device such as a UAV, its limited compu...

AI Technical Summary

Problems solved by technology

The first is that the current popular target recognition network often consumes a lot of parameters and computing resources, so it is a fatal flaw for various embedded devices, especially drones.

In addition, due to limited computing resources, it takes a long time for the target recognition network to process a frame of image, so it is difficult for UAVs to have real-time detection capabilities

Images