Pavement crack identification method based on transposed neural network interlayer feature fusion

Abstract

The invention discloses a pavement crack identification method based on transposed neural network interlayer feature fusion, and relates to the field of computer vision and road engineering. Accordingto the method provided by the invention, a strategy of fusing features of an intermediate layer of a classification network to optimize the segmentation result so as to compensate the receptive fieldlimitation is provided according to the information contained in the intermediate layer of the classification network and the receptive field limitation of the segmentation network. Upsampling of theintermediate layer is carried out through fractional step convolution, three independent transposed convolutional networks are constructed based on output feature maps of the three pooling layers after the classification network, and the three pooling layers are fused through layer-by-layer transposed convolution to construct a transposed convolutional network CNN-T with the three transposed convolutional layers. Results show that on the basis that the detail features of the segmentation result are basically reserved, the continuity of the crack features is perfected to a great extent by theCNN-T.

Description

technical field

[0001] The invention belongs to the fields of computer vision and road engineering, and in particular relates to a pavement crack recognition method based on interlayer feature fusion of a transposed neural network. Background technique

[0002] In order to describe the crack information more accurately, pixel-accurate image segmentation methods are applied in the field of disease image recognition. In order to achieve end-to-end network input / output, there are generally two options: a network structure without downsampling throughout the process, or downsampling-upsampling Combined network structure. The former represents the CrackNet series model, which cancels the pooling layer to ensure that the entire size of the feature map remains unchanged. This strategy has high requirements for computing time and computing power while ensuring accuracy. Representatives of the latter are Fully-Connected Network (FCN) and U-Net. Its structure has undergone the proc...

Images