Multi-label image preprocessing method for gesture recognition based on deep learning

Abstract

The invention discloses a multi-label image preprocessing method for gesture recognition based on deep learning. The steps of the present invention are as follows: 1. obtain the coordinates of a plurality of marked feature points; 2. select the maximum and minimum values ​​of x in the image plane coordinate system in the coordinates, the maximum and minimum values ​​of y, and confirm the coordinate point A (x min ,y min ) and coordinate point B(x max ,y max ); 3. Use the coordinates A and B to locate the square area P where the target label is valid in the image, and at the same time, a margin should be left at the coordinate points on the edge of the area P according to the rules, so as to obtain the expanded square area P 1 , and for the region P 1 The length and width of the update; 4. Compare the area P 1 length and width to get a new square area P 2 ; 5. Calculate the coordinates of the feature points in the clipping image as labels of the clipping image. The present invention cuts out a square image containing a target from an original image, adds as few channels as possible, reduces background redundancy, and retains target features.

Description

technical field [0001] The invention relates to an image preprocessing method for gesture recognition based on deep learning, which is suitable for feature points with complex backgrounds and multiple labels as two-dimensional coordinates of the recognized target. Background technique [0002] In order to make the loss value of the neural network converge better and faster, and to obtain a model with an excellent recognition rate, the trained 3D images will undergo a series of preprocessing before being input into the network. At present, no matter whether researchers use deep learning for classification tasks or regression tasks, they will directly scale the length and width of the original image of the dataset to the same size, or add 0 channels on the short side to ensure that the original image’s aspect ratio remains unchanged. Make the length and width consistent, which is the normalization of the scale. But the consequence of this is that the former compresses the asp...

AI Technical Summary

Problems solved by technology

But the consequence of this is that the former compresses the aspect ratio of the original image, causing the detected target object in the image to deform, losing texture features, and reducing the accuracy rate; although the latter retains texture features, it will add many of the same The channel of the pixel value, which does not belong to the information in the image, increases the redundancy of the image, so it will also have a great impact on the accuracy

[0003] At the same time, when the proportion of the target in the image is small and the background is complex, there is too much useless information in the picture, which is not conducive to the extraction of target features by the neural network. The background is redundant and noisy, and the neural network will It is not easy to converge during training, the training time is longer, it will burden the server GPU, and the accuracy of the model will also be affected

Images