An improved RoIAgign region feature aggregation algorithm

Abstract

The invention discloses an improved RoIAgign region feature aggregation algorithm, and the method comprises the following steps: enabling the geometric centers of a source image and a target image tobe aligned, and enabling the alignment method to adopt improved central point alignment; Then, performing bilinear interpolation calculation, and converting floating point operation into integer operation in order to improve the calculation efficiency and the calculation precision; According to the invention, a RoIAgign region feature aggregation method is optimized; central point coordinates of the source image and the target image are overlapped; all pixels of the source image are fully and uniformly utilized in the mapping process; Compared with the prior art, the method has the advantagesthat the region information loss caused by feature map pooling is avoided, the target image interpolation effect is better, the calculation amount is greatly reduced by converting the floating point operation into the integer operation, and the calculation efficiency of the shift operation is much higher than that of the common multiplication and division operation, so that the calculation efficiency is improved, and the hardware requirement is reduced.

Description

technical field [0001] The invention belongs to the field of regional feature-based aggregation method algorithms, and in particular relates to an improved RoIAlign regional feature aggregation algorithm. Background technique [0002] ROI Align is a regional feature aggregation method proposed in the Mask-RCNN paper, which solves the problem of mis-alignment caused by two quantizations in the ROI Pooling operation. Experiments show that replacing ROI Pooling with ROI Align in the detection task can improve the accuracy of the detection model. [0003] However, in the prior art, when mapping the entire target image to the source image, the pixels used are concentrated in the upper left of the image, and the pixels in the lower right part are not used, and with the increase of the interpolation ratio, the uneven phenomenon is more serious. Obviously, and due to the use of floating-point operations, the calculation takes a long time and requires high equipment. Therefore, a ne...

AI Technical Summary

Problems solved by technology

[0003] However, in the prior art, when mapping the entire target image to the source image, the pixels used are concentrated in the upper left of the image, and the pixels in the lower right part are not used, and as the interpolation ratio increases, the unevenness becomes more Obviously, and due to the use of floating-point operations, the calculation takes a long time and requires high equipment requirements. Therefore, a new technology is urgently needed to optimize the RoIAlign regional feature aggregation algorithm.

Images