Pooling feature map processing method, target detection method, target detection system, pooling feature map processing device and medium

Abstract

The invention discloses a pooling feature map processing method, an image target detection method and, a pooling feature map processing device and a medium. The pooling feature map processing method comprises the following steps: respectively carrying out convolution activation processing on a pooling feature map twice, carrying out channel splicing processing on the result of the convolution activation processing twice; and outputting the maximum value of the corresponding position of each channel in the channel splicing processing result. A; and performing activation processing on the output result, performing element-by-element multiplication calculation on the result of the activation processing and the pooling feature map, performing element-by-element addition calculation on the result of the element-by-element multiplication calculation and the pooling feature map, and the like. According to the method, the pixel value information of the important region in the pooling feature map can be amplified, so that the pixel loss generated by the Faster RCNN algorithm in the region pooling process is made up, and the target detection accuracy of the Faster RCNN algorithm is improved. The method is widely applied to the technical field of image recognition.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a pooling feature map processing method, a target detection method, a system, a device and a medium. Background technique [0002] In the field of image recognition technology such as face recognition and license plate monitoring, the Faster RCNN algorithm is widely used to extract the target object in the image to be processed. Faster RCNN algorithm such as figure 1 As shown, the feature extraction network T constructed by using convolutional neural networks such as ZF, VGG16 or RES101 extracts the overall feature map P from the image to be processed, and then the overall feature map P is input into the PRN network, and processed by the RPN network After generating multiple candidate regions (such as 2000), and then select some candidate regions (such as 300) through the NMS algorithm, and map the selected candidate regions to the overall feature map P to obtain the c...

AI Technical Summary

Problems solved by technology

[0003] The pooling algorithm is used in the execution of Faster RCNN, and the quantization operation in it will cause a certain degree of distortion in the feature map of the candidate area, resulting in loss of detail information and discernible Accuracy decreases, thereby reducing the accuracy of the recognition results of the target object

Images