Generative adversarial network-based lightweight efficient target segmentation and counting method

Abstract

The invention relates to the technical field of video image processing, in particular to a lightweight efficient target segmentation and counting method based on a generative adversarial network. According to the invention, the folded nearest neighbor surpassing up-sampling method is provided at the decoder stage, so that the calculation amount is greatly reduced, the network operation is accelerated, the network operation efficiency is improved, and the network structure is greatly optimized; in the predictor stage, an independent predictor is set for each task so as to meet unique requirements of different tasks; the discriminator is subjected to lightweight design, the model structure is simplified, and the training process is accelerated; a target quantity statistics task based on the density map is split into two tasks of quantity prediction and position prediction, so that the learning difficulty is reduced, and the usable range of a data set during pre-training is expanded; according to the training method based on the generative adversarial network, the problems of low training speed, low efficiency and complex structure of a multi-task generator used during segmentation and counting of a target image are solved.

Description

technical field [0001] The invention relates to the technical field of video image processing, in particular to a lightweight and efficient target segmentation and counting method based on a generative confrontation network. Background technique [0002] At present, the counting of targets mainly adopts the method of density map; this method can display the position distribution of the target through the density map and obtain the total number of targets by calculating the value of the density map. Although this method achieves the counting and distribution of targets at the same time, this method puts forward higher requirements for the complexity of the network and the collection of data sets. When using the density map method, the data set requires labeling each target point in the image to obtain a point map with precise coordinate positions. Such labeling work is time-consuming and laborious; and then the point map is calculated as a density map through mathematical cal...

AI Technical Summary

Problems solved by technology

Although this training method is an advanced training method for most neural networks, it usually requires the design of a unique multi-column network model to deal with multiple tasks and more complex loss functions when training a multi-task model. to coordinate multi-task goals, and the training speed is usually slower, and it takes longer to complete multi-task training

Images