Image processing method, device, and computer-readable storage medium

Abstract

The present application provides an image processing method and device. The method includes: dividing the image to be detected into Q non-overlapping grids, wherein Q is an integer greater than 1, and the p-th grid in the Q grids belongs to K(p) regions, wherein, p=1, 2,..., Q, K(p) is an integer greater than 1, and the K(p) regions occupy different regions on the image to be detected; for any region in the K(p) regions The sub-feature map is intercepted on the high-level feature map, and the targets to be detected in the K(p) regions are classified and located according to the feature map. This application can divide the same position into regions of different scales, so that sub-feature maps of different scales corresponding to the position can be extracted on the high-level feature map, making the sub-feature map extracted at this position more comprehensive and accurate. Therefore, the quality of the generated target candidate area can be improved, and the performance of the target detection system can be improved.

Description

technical field [0001] The present application relates to the field of image processing, and more specifically, to an image processing method, device, and computer-readable storage medium. Background technique [0002] Object detection is a basic task in the field of computer vision, which aims to identify and locate objects of interest in images or videos, and has a very wide range of applications in the industry. Object detection needs to identify the objects that appear in the image and belong to the predefined categories, and accurately locate each object in the form of a detection box. Therefore, the target detection task mainly faces two challenges: [0003] (1) It is necessary to distinguish the target from the complex background; [0004] (2) The candidate boxes of the target need to be tuned to make them closer to the real frame of the target to obtain higher accuracy. [0005] With the advent of deep learning, most recent object detection systems are based on de...

AI Technical Summary

Problems solved by technology

However, there are some problems in the target discrimination of RPN network

On the one hand, RPN only uses the "target / non-target" two-class label for training, and the learned classifier lacks the ability to discriminate complex features, and it is easy to misreport the background as the target during prediction.

On the other hand, RPN uses a 3×3 sliding window to extract features on the high-level features of the convolutional neural network (CNN), which will cause the receptive field of the larger target to be much smaller than the scale of the target itself, resulting in the loss of target information. lost

[0007] In view of the above problems, some research works try to directly change the two-class classification layer of RPN into a multi-class classification layer, but the effect of predicting multi-class targets based on the features extracted by a fixed 3×3 sliding window is not good, so the generated The quality of object proposals is not high, resulting in poor performance of object detection systems

Images