US 12,639,924 B2 · Three-stage modularized convolutional neural network for rapidly classifying concrete cracks

Abstract

Disclosed is a three-stage modular convolutional neural network (CNN), referred to as Stairnet, for efficient classification of concrete cracks in images. Unlike conventional CNNs that employ repetitive layer structures, Stairnet is organized into three distinct stages—Stair1, Stair2, and Stair3—each with unique architectural characteristics. Stair1 is composed solely of convolution layers; Stair2 introduces a higher layer count and incorporates channel-split and shuffle operations; and Stair3 utilizes larger expansion factors and kernel sizes along with attention mechanisms. Variations in Stair1 and Stair2 enable structural adaptability based on model hyperparameters. Compared to general-purpose CNNs designed for thousands of classes, Stairnet achieves a compact model size, faster training, and high accuracy in concrete crack classification.

A three-stage modular convolutional neural network (CNN) for rapid classification of concrete cracks in images, comprising:

an input layer configured to receive an input image;
a shallow-layer feature extraction module, designated as Stair1, operatively connected to the input layer, wherein Stair1 comprises two basic block_1 structural blocks, each basic block_1 comprising an inverted residual block with convolutions (Convs) and having an expansion factor configurable as an integer during model initialization, wherein:
when the expansion factor is not 1, the basic block_1 consists of a first 3×3 Conv layer and a second 1×1 Conv layer;
and
when the expansion factor is 1, the basic block_1 consists of a single 3×3 Conv layer;
a first convolutional block attention module (CBAM) operatively connected to the shallow-layer feature extraction module (Stair1);
a mid-layer feature extraction module, designated as Stair2, operatively connected to the first CBAM, wherein Stair2 comprises a plurality of basic block_2 structural blocks each having a stride configurable to 1 or 2 during model initialization, wherein:
when configured for a stride of 1, the basic block_2 comprises a dual-branch structure having:
(i) a partitioner unit configured to partition an input feature map of basic block_2 (stride=1) into a first portion and a second portion;
(ii) a first processing branch (stride=1), coupled to the partitioner unit, configured to process the first portion, the first processing branch comprising a first 1×1 Conv layer, a 3×3 depthwise separable convolutional (DConv) layer, and a second 1×1 Conv layer connected in series;
(iii) a second processing branch (stride=1), coupled to the partitioner unit, configured to pass through the second portion without processing;
(iv) a concatenator unit (stride=1), coupled to outputs of the first and second processing branches (stride=1), configured to concatenate an output of the first processing branch and an output of the second processing branch;
and
(v) a first channel shuffle unit, coupled to an output of the concatenator unit (stride=1), configured to apply a channel shuffle operation to the concatenated output;
when configured for a stride of 2, the basic block_2 comprises a triple-branch structure having:
(i) a replicator unit configured to replicate an input feature map of basic block_2 (stride=2) into a first copy, a second copy, and a third copy;
(ii) a first processing branch (stride=2), coupled to the replicator unit, configured to process the first copy, the first processing branch comprising a first 1×1 Conv layer, a 5×5 DConv layer, and a second 1×1 Conv layer connected in series;
(iii) a second processing branch (stride=2), coupled to the replicator unit, configured to process the second copy, the second processing branch comprising a 5×5 DConv layer and a 1×1 Conv layer connected in series;
(iv) a third processing branch (stride=2), coupled to the replicator unit, configured to process the third copy, the third processing branch comprising a 3×3 max pooling layer and a 1×1 Conv layer connected in series;
(v) a concatenator unit, coupled to outputs of the first, second, and third processing branches (stride=2), configured to concatenate the outputs of the first, second, and third processing paths (stride=2);
and
(vi) a second channel shuffle unit, coupled to an output of the concatenator unit (stride=2), configured to apply a channel shuffle operation to the concatenated output;
a second CBAM operatively connected to the mid-laver feature extraction module (Stair2);
a deep-layer feature extraction module, designated as Stair3, operatively connected to the second CBAM, wherein Stair3 comprises an inverted residual structure with an efficient channel attention (ECA) module, the inverted residual structure in Stair 3 sequentially comprising:
a first 1×1 Conv layer, a 5×5 DConv layer, the ECA module, and a second 1×1 Conv layer;
and
a fully connected layer operatively connected to the deep-layer feature extraction module (Stair3), configured to output a crack classification result.

24 more elements

Three-stage modularized convolutional neural network for rapidly classifying concrete cracks

Life of this patent

Three-stage modularized convolutional neural network for rapidly classifying concrete cracks

Life of this patent

Abstract

Claims8