Road sign detection method based on region full convolution neural network

Abstract

The invention discloses a road sign detection method based on a region full convolution neural network. A road sign is detected using a full convolution neural network and a region recommendation network. The robustness of road sign detection is improved. The method comprises the following steps: S1, using a convolution layer to extract the features of an image; S2, using a region recommendation network to extract candidate regions; and S3, using a region-based full convolution neural network to classify the candidate regions; and S4, outputting a road sign detection result.

Description

technical field [0001] The invention relates to computer-based image processing technology, in particular to a road sign recognition method based on a region-based fully convolutional neural network. Background technique [0002] In daily traffic driving, traffic signs play an important role, and the correct automatic detection of traffic signs has potential application value. The deep network of object detection is divided into two mainstreams according to the region of interest (RoI) pooling layer: the fully convolutional sub-network that shares computation (each sub-network has nothing to do with RoI) and the sub-network that does not share computation and acts on its own RoI. Engineering classification structures (such as Alexnet and VGG Nets) lead to such splitting, while engineering image classification structures are designed as two sub-networks - 1 suffix, 1 convolutional sub-network of spatial pooling layer and multiple fully connected layer. Therefore, the last s...

AI Technical Summary

Problems solved by technology

However, experience in object detection work shows that the detection performance of this solution is much worse than the classification performance of this network

However, the faster R-CNN detector unnaturally inserts the RoI pooling layer between the two convolutional layers, so that the deeper sub-network acting on each RoI has higher accuracy, but the calculation of each RoI is not shared, so the speed is slow

[0005] In the existing methods, the detection accuracy based on the full convolutional neural network method is not high, and the detection speed of the RoI-based R-CNN is slow. Therefore, the present invention discloses a street sign detection method based on a region-based fully convolutional neural network , the present invention adopts the popular object detection strategy, the first step is to use the volume base layer to extract the features of the image; the second step is to use the region proposal network to extract candidate regions; the second step is to use the region-based fully convolutional neural network to extract candidate regions Classify the area; finally output the detection result of the road sign

Images