A crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion

Abstract

The application provides a crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, first inputting an input picture into a parallel hole convolution module to obtain a large receptive field multi-scale fusion feature map; then inputting into an encoding layer, the encoding layer being composed of a convolution layer, a self-optimization module and a pooling layer, performing one-time information screening optimization on the encoding layer feature map through the self-optimization module to obtain an encoding layer feature map; inputting into a decoding layer again to obtain decoding layer feature maps at different stages; fusing the feature maps corresponding to the encoding layer and the decoding layer to obtain an intermediate fusion layer feature map, and inputting into four weighted fusion modules to obtain weighted fusion feature maps and layered independent prediction maps; finally fusing and reducing the dimension of the layered independent prediction maps to obtain a final prediction map. The application fully utilizes semantic information at different stages and feature fusion at different layers to obtain an optimized prediction result and promote the progress of crack detection automation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crack detection deep learning, and in particular but not exclusively to a crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion. BACKGROUND

[0002] The deep learning method is continuously optimized by using a data set training to adapt to the task to be solved, and the method is introduced into the crack detection problem to achieve better results than the traditional image processing algorithm. At present, the main solution to the crack detection work in the industry is still the traditional manual method, which consumes a large amount of resources. The semantic segmentation algorithm based on deep learning in the image processing field can realize pixel-level crack labeling, and through the collection of the photographed crack data set for model training, the automatic crack detection can be realized. The deep learning model in the field of semantic segmentation is developing rapidly, and many model design methods suitable for this task have been produced, such as coding and decoding structure, feature fusion, hollow convolution, attention mechanism, etc. From these angles, a new method can be designed to obtain a better trainable crack detection model.

[0003] In the classical semantic segmentation algorithm, the input picture is usually directly coded and decoded: the picture is first down-sampled through successive convolution and pooling processing, and then the feature map output by the encoding layer is up-sampled through the decoding layer, and finally the prediction map with the same size as the input picture is obtained. There are many problems in such a process, the encoding layer feature map is rough, and the encoding layer feature and the decoding layer feature are not fully utilized. Moreover, using convolution and pooling to obtain the receptive field has defects, and the subsequent obtained receptive field is based on the local information feature map captured by the bottom convolution, and the global information is missing. In order to solve these problems, a feature map self-optimization module and a multi-layer feature step-by-step weighted fusion mechanism are proposed to optimize the detection model.

[0004] Therefore, it is necessary to provide a new detection method to solve at least part of the above problems. SUMMARY

[0005] In view of one or more problems in the prior art, the present application proposes a crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, which adopts parallel hollow convolution, self-optimization and multi-layer feature weighted fusion module, predicts according to the input crack picture, and outputs pixel-level crack prediction map. On the basis of the basic segnet network, the feature map self-optimization and the multi-layer feature step-by-step weighted fusion are realized to fully utilize the coding and decoding feature maps, so as to obtain a better trainable neural network model for crack detection, promote the progress of crack detection automation, and solve the problems of the coding and decoding model structure that cannot fully utilize the multi-layer features and the small feature roughness of the coding feature map receptive field.

[0006] The technical solution for achieving the object of the present application is:

[0007] A crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, comprising the following steps:

[0008] Step 1: input the picture into a parallel hole convolution module to obtain large receptive field multi-scale fusion feature maps;

[0009] Step 2: input the fusion feature maps into n encoding layers in sequence, each encoding layer inputs the optimized encoding feature maps into the next encoding layer to obtain n encoding layer feature maps of different scales; each encoding layer comprises a plurality of convolution layers, a feature map self-optimization module and a pooling layer, the feature map self-optimization module is used for information screening and optimization of the generated feature map according to its own encoding information, and n is a positive integer;

[0010] Step 3: input the encoding layer feature maps into n decoding layers in sequence to obtain n decoding layer feature maps of different scales;

[0011] Step 4: input the encoding layer feature maps and the decoding layer feature maps into a feature fusion module, and use a channel splicing method to fuse the corresponding encoding layer feature maps and the decoding layer feature maps to obtain n intermediate fusion layer feature maps of different scales;

[0012] Step 5: use a plurality of weighted fusion modules to perform hierarchical weighted fusion on the intermediate fusion layer feature maps, distribute weights from high-level intermediate fusion layer features to low-level features step by step and perform fusion to obtain hierarchical fusion feature maps, and generate an independent prediction map each time using each fused hierarchical feature to generate n-1 layered independent prediction maps;

[0013] Step 6: fuse and reduce the dimension of the n-1 layered independent prediction maps to obtain a final prediction map.

[0014] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the parallel hole convolution module in step 1 comprises a plurality of controllable hole rate hole convolution modules, a splicing module and a convolution module, the three hole convolution modules are connected in parallel and then connected in sequence to the splicing module and the convolution module, and the parallel hole convolution module can obtain feature maps of different scales by adjusting the hole rate.

[0015] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the feature map self-optimization module in step 2 comprises a sigmoid function module, a multiplier and an adder, the input end of the sigmoid function module is connected to the input data, the output end of the sigmoid function module is connected to the first input end of the multiplier, the multiplier and the adder are connected in sequence, the second input end of the multiplier is connected to the learnable parameter α, and the second input end of the adder is connected to the input data.

[0016] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the feature map self-optimization module in step 2 comprises a sigmoid function module, a multiplier and an adder, the input end of the sigmoid function module is connected to the input data, the output end of the sigmoid function module is connected to the first input end of the multiplier, the multiplier and the adder are connected in sequence, the second input end of the multiplier is connected to the learnable parameter α, and the second input end of the adder is connected to the input data.

[0017] enfmp′ i =sigmoid(enfmp i )*α+enfmp i (i=1,2,…,n)

[0018] wherein enfmp i is the original encoding feature map, α is the self-adaptive learnable parameter, and sigmoid() is an activation function.

[0019] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the encoding feature map input to the next encoding layer in step 2 is:

[0020] enfmp i+1 =P(enfmp′ i )(i=1,2,…,n-1)

[0021] wherein P() represents pooling, and enfmp′ i represents the encoding feature map optimized by the feature map self-optimization module.

[0022] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the feature fusion module in step 4 comprises a splicing module and a 1×1 convolution module, the input ends of the splicing module are connected to the output data of the encoding layer and the decoding layer respectively, the output end of the splicing module is connected to the 1×1 convolution module, and the output end of the 1×1 convolution module outputs the intermediate fusion layer feature map.

[0023] Further, in the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, the intermediate fusion layer feature map in step 4 is:

[0024] hfmp i =conv(concat(enfmp i′ ,defmp i ))(i=1,2,3,4,5)

[0025] wherein, conv() represents 1x1 convolution used to reduce dimension of feature map, concat() represents channel concatenation operation, defmp i represents decoding layer feature map, enfmp i ′ represents encoding feature map optimized by feature map self-optimization module.

[0026] Further, the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application, step 5, using a hierarchical weighted fusion module to perform hierarchical weighted fusion on the intermediate fusion layer feature map specifically includes:

[0027] Step 5-1: extracting semantic information of the intermediate fusion layer feature map of the high level to form the i-1 layer weight feature map Qfmp i-1 :

[0028] Qfmp i-1 =res((res(conv(hfmp i ))*(res(sigmoid(conv(hfmp i ))) T )*res(hfmp i-1 ))

[0029] wherein, res() represents feature map deformation operation, conv() represents 1x1 convolution, defmp i represents decoding layer feature map, sigmoid represents

[0030] , i=1,2,…,n;

[0031] Step 5-2: weighting the weight feature map Qfmp i-1 to the intermediate fusion layer feature map hfmp i-1 of the lower layer and fusing to obtain the fused lower layer hierarchical fusion feature map fhfmp i-1 :

[0032] fhfmp i-1 =(hfmp i-2 +Qfmp i-2 )+hfmp i-1

[0033] Step 5-3: obtaining the layered independent prediction map pre i based on the hierarchical fusion feature map fhfmp i :

[0034] pre i =conv(fhfmp i )(i=1,2,…,n)

[0035] Wherein, conv() represents 1x1 convolution.

[0036] Further, the crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application, the last prediction map pre in step 6 is:

[0037] pre=sigmoid(convp(concat(pre i ))(i=1,2,…,n)

[0038] Wherein, convp() represents 3x3 convolution, which is used to generate the prediction map, concat() represents the channel splicing operation, pre i is a hierarchical independent prediction map.

[0039] Compared with the prior art, the application has the following technical effects:

[0040] 1. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application realizes efficient feature reuse on the basis of the basic coding and decoding network, and further optimizes the prediction effect.

[0041] 2. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application introduces a parallel hollow convolution module and a feature map self-optimization module, which can optimize the feature map from two angles of receptive field and spatial details, and can obtain more accurate encoding layer feature map of encoding information.

[0042] 3. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application adds step-by-step weighting design to the basic feature fusion structure, introduces high-level semantic information to feature fusion, optimizes the fusion effect, and provides more level information basis for the final crack detection. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the application, together with the description, to explain the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0044] Figure 1 The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application is shown in the flowchart.

[0045] Figure 2 The basic coding and decoding SegNet network structure schematic diagram is shown.

[0046] Figure 3 A schematic diagram of a crack detection network structure based on feature map self-optimization and multi-layer feature step-by-step weighted fusion of the application is shown.

[0047] Figure 4 A schematic diagram of a parallel hole convolution module of the application is shown.

[0048] Figure 5 A schematic diagram of a self-optimization module of the application is shown.

[0049] Figure 6 A schematic diagram of a feature fusion module of the application is shown.

[0050] Figure 7 A schematic diagram of a weighted fusion module of the application is shown. DETAILED DESCRIPTION

[0051] In order to further understand the application, the preferred embodiments of the application are described below in combination with the embodiments, but it should be understood that the description is only for further illustrating the features and advantages of the application, and is not a limitation on the claims of the application.

[0052] The description of this part is only for typical embodiments, and the application is not limited to the scope described in the embodiments. The combination of different embodiments, the mutual replacement of some technical features in different embodiments, and the mutual replacement of some technical features in the same or similar prior art means are also within the scope of the description and protection of the application.

[0053] The application is built on the basis of a basic coding and decoding SegNet network (as shown in Figure 2 ), and a structure (as shown in Figure 3 ) for crack detection is built, a coding and decoding structure design is adopted, the coding stage is further optimized by adding a parallel hole convolution module and a self-optimization module, corresponding layer feature fusion and layer-by-layer weighted feature fusion are added between the coding and decoding stages, and the prediction stage is optimized. From two major directions, the detection ability of the entire model is optimized.

[0054] Embodiment 1

[0055] According to an aspect of the application, a crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, as shown in Figure 1 , comprises the following steps:

[0056] Step 1: input the picture into a parallel hole convolution module (Para-dconv) to obtain large receptive field multi-scale fusion feature maps, and the structure of the parallel hole convolution module is shown in Figure 4The parallel hollow convolution module includes three hollow convolution deconv modules with controllable hollow rates, and the output is obtained through a concatenation operation concat and a convolution operation conv. By adjusting the hollow rate, feature maps of different scales can be obtained.

[0057] Step 2: refer to Figure 3 The left half of the structure is the encoding stage, which encodes the picture. Specifically, the fused feature map is sequentially input into five encoding layers, and each encoding layer inputs the optimized encoding feature map into the next encoding layer to obtain five encoding layer feature maps of different scales. Each encoding layer includes a plurality of convolution layers, a feature map self-optimization module (self-opt) and a pooling layer. The structure of the self-optimization module is shown in Figure 5 The input is multiplied by a learnable parameter a after passing through a sigmoid function, and finally added to the original input to obtain the output. The feature map self-optimization module is used to filter and optimize the generated feature map according to its own encoding information, which can optimize the feature map in space. The optimized encoding feature map is:

[0058] enfmp′ i = sigmoid(enfmp i )*a + enfmp i (i = 1, 2,..., 5)

[0059] where enfmp i is the original encoding feature map, a is an adaptive learnable parameter, and sigmoid() is an activation function:

[0060] Then the encoding feature map input into the next encoding layer is:

[0061] enfmp i+1 = P(enfmp′ i )(i = 1, 2,..., 4)

[0062] where P() represents pooling, and enfmp′ i represents the optimized encoding feature map after the feature map self-optimization module.

[0063] Step 3: refer to Figure 3 The right half of the structure is the decoding stage, which is designed symmetrically with the encoding stage. Each corresponding decoding layer up-samples the encoding layer feature map, and finally restores to the input picture size. Specifically, the encoding layer feature map is sequentially input into five decoding layers to obtain five decoding layer feature maps of different scales.

[0064] Step 4: refer to Figure 3The corresponding layer fusion is performed on the coded feature map, specifically: the corresponding coded layer feature map and the decoded layer feature map with the same size are input into a feature fusion module (fuse), and the structure of the feature fusion module is referred to in Figure 6 The inputs of the coded layer and the decoded layer are spliced on the channel through a concat operation, and then the output is obtained through a 1x1 convolution. The corresponding coded layer feature map and the decoded layer feature map are fused using a channel splicing method, and then dimension reduction is performed to obtain five intermediate fusion layer feature maps with different scales:

[0065] hfmp i = conv(concat(enfmp i , defmp i ))(i = 1, 2, …, 5)

[0066] Wherein, conv() represents a 1x1 convolution used for dimension reduction of the feature map, concat() represents a channel splicing operation, defmp i represents a decoded layer feature map, enfmp i represents a coded feature map optimized by a feature map self-optimization module.

[0067] Step 5: refer to Figure 3 The step-by-step weighted fusion is performed on the coded and decoded feature maps, specifically: a plurality of weighted fusion modules (weighted fuse) are used to perform hierarchical weighted fusion on the intermediate fusion layer feature maps, and the structure of the weighted fusion module is referred to in Figure 7 The deep layer input is subjected to a reshape and a sigmoid function, and then multiplied to obtain a weight feature map. The shallow layer input is subjected to a reshape operation and multiplied with the weight feature map, and then reshaped to the original input shape. The result is added to the deep layer input after upsampling, and the output is obtained. By distributing the weight of the high-level intermediate fusion layer feature to the low-level step by step and performing fusion, the hierarchical fusion feature map is obtained. At the same time, each fused hierarchical feature generates an independent prediction map, and four hierarchical independent prediction maps are generated. The specific steps include:

[0068] Step 5-1: extract the semantic information of the high-level intermediate fusion layer feature map to form the i-1 layer weight feature map Qfmp i-1 :

[0069] Qfmp i-1 = res((res(conv(hfmp i ))*(res(sigmoid(conv(hfmp i ))) T )*res(hfmp i-1 ))

[0070] wherein res() represents a feature map deformation operation, conv() represents a 1x1 convolution, defmp i represents a decoded layer feature map, and sigmoid represents an activation function:

[0071] Step 5-2: The weight feature map Qfmp i-1 is assigned to the low-level intermediate fusion layer feature map hfmp i-1 and fused to obtain the fused low-level hierarchical fusion feature map fhfmp i-1 :

[0072] fhfmp i-1 = (hfmp i-2 + Qfmp i-2 ) + hfmp i-1

[0073] Step 5-3: Based on the hierarchical fusion feature map fhfmp i , a layered independent prediction map pre i is obtained:

[0074] pre i = conv(fhfmp i )(i = 1, 2, …, 5)

[0075] wherein conv() represents a 1x1 convolution.

[0076] Step 6: The four layered independent prediction maps are input into a pre module, the independent prediction maps are spliced by channels, and a final prediction map pre is obtained through 3x3 convolution and fusion dimension reduction:

[0077] pre = sigmoid(convp(concat(pre i ))(i = 1, 2, …, n)

[0078] wherein convp() represents a 3x3 convolution for generating a prediction map, concat() represents a channel splicing operation, and pre i is a layered independent prediction map.

[0079] Embodiment 2

[0080] The application adopts a codec structure to design a crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, and the method is specifically as follows:

[0081] Stage 1, referring to FIG. 1, the left half is an encoding stage, and the input picture is processed by encoding. Figure 2 Stage 1.1, referring to FIG. 2, the left half is an encoding stage, and the input picture is processed by encoding.

[0082] Figure 2 ​, the input picture first passes through a parallel dilated convolution module (Para-dconv) to capture multi-scale large receptive field feature maps, and the structure of the parallel dilated convolution module is shown in FIG. 1 Figure 3 .

[0083] Stage 1.2, see FIG. 2 Figure 2 , after each convolution layer in the encoding stage, a self-optimization module (self-opt) is used, and the structure of the self-optimization module is shown in FIG. 2 Figure 4 , which can realize feature map optimization in space.

[0084] Stage 2, see FIG. 3 Figure 2 , the right half is the decoding stage, which is designed symmetrically with the encoding stage, and each corresponding decoding layer up-samples the encoding layer feature map, and finally restores to the input picture size.

[0085] Stage 3, see FIG. 4 Figure 2 , feature fusion is performed on the encoding and decoding feature maps, which is divided into two sub-stages: corresponding layer fusion (fuse) and weighted fusion (weighted fuse).

[0086] Stage 3.1, see FIG. 5 Figure 5 , the same size corresponding encoding and decoding feature maps are fused by channel splicing and then dimensionally reduced to obtain hierarchical fusion feature maps.

[0087] Stage 3.2, see FIG. 6 Figure 6 , the hierarchical feature maps are taken as input, the semantic information of the high-order feature maps is first extracted to become a weight feature map, then the weight feature map is assigned to the low-order feature map, and then the fusion is performed.

[0088] Stage 4, see FIG. 7 Figure 2 , the prediction is divided into two stages:

[0089] Stage 4.1, see FIG. 8 Figure 6 , the weighted fusion module (weighted fuse) not only outputs the fusion feature map, but also reduces the dimension of the fusion feature map to generate an independent prediction.

[0090] Stage 4.2, see FIG. 9 Figure 2 , the independent prediction of each layer is input into the pre module, the independent prediction maps are spliced by channels, and a 3x3 convolution is used for final prediction.

[0091] The description and applications of the present application are illustrative, and not intended to limit the scope of the present application. The related descriptions of effects or advantages etc. involved in the specification are not intended to limit the scope of the present application, and are only intended to describe the related descriptions of effects or advantages etc. in the specification. Variations and changes may be made to the disclosed embodiments, and it is intended that the application encompass variations and modifications. It is intended that the application not be limited to the embodiments disclosed in the specification, but include all embodiments that would normally occur to one of ordinary skill in the art. It is intended that the application not be limited to the disclosed embodiments, but include all embodiments that would normally occur to one of ordinary skill in the art. The person skilled in the art should understand that the present application can be implemented in other forms, structures, arrangements, proportions, and with other components, materials and parts without departing from the spirit or essential characteristics of the present application. Other variations and changes may be made to the disclosed embodiments without departing from the scope and spirit of the application.

Claims

1. A crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighted fusion, characterized in that, The following steps are involved: Step 1: Input the image into the parallel dilated convolution module to obtain a large receptive field multi-scale fusion feature map; Step 2: Input the fusion feature map into n coding layers in sequence, and each coding layer inputs the optimized coding feature map into the next coding layer to obtain n coding layer feature maps of different scales; each coding layer includes several convolution layers, feature map self-optimization modules and pooling layers, and the feature map self-optimization module is used to perform information screening and optimization on the generated feature map according to its own coding information, and n is a positive integer; the feature map self-optimization module includes a sigmoid function module, a multiplier and an adder, the input end of the sigmoid function module is connected to the input data, the output end of the sigmoid function module is connected to the first input end of the multiplier, the multiplier and the adder are connected in sequence, the second input end of the multiplier is connected to the learnable parameter α, and the second input end of the adder is connected to the input data; the optimized coding feature map is: enfmp′ i = sigmoid(enfmp i )*a + enfmp i (i = 1, 2,..., n) wherein enfmp i is the original encoded feature map, a is an adaptive learnable parameter, and sigmoid() is an activation function: Step 3: Input the encoding layer feature map into n decoding layers in sequence to obtain n decoding layer feature maps of different scales; Step 4: Input the encoding layer feature map and the decoding layer feature map into the feature fusion module, and use the channel splicing method to fuse the corresponding encoding layer feature map and decoding layer feature map to obtain n intermediate fusion layer feature maps of different scales; Step 5: Use several weighted fusion modules to perform hierarchical weighted fusion on the intermediate fusion layer feature maps, assign weights to the lower layers of the high-level intermediate fusion layer features step by step and fuse them to obtain a hierarchical fusion feature map. At the same time, use each fused hierarchical feature to generate an independent prediction map, generating n-1 hierarchical independent prediction maps. Step 6: Fuse and reduce the dimension of n-1 hierarchical independent prediction maps to obtain the final prediction map.

2. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1, characterized in that, The parallel dilated convolution module in step 1 includes three dilated convolution modules with controllable dilated rates, a splicing module, and a convolution module. The three dilated convolution modules are connected in parallel and then connected to the splicing module and the convolution module in sequence. The parallel dilated convolution module can obtain feature maps of different scales by adjusting the dilated rate.

3. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1, characterized in that, The encoded feature map input to the next encoding layer in step 2 is: enfmp i+1 = P(enfmp′ i )(i = 1, 2,..., n - 1) wherein P() represents pooling, enfmp' i denotes the encoding feature map optimized by the feature map self-optimization module.

4. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1, characterized in that, The feature fusion module in step 4 includes a splicing module and a 1×1 convolution module. The input end of the splicing module is connected to the output data of the encoding layer and the decoding layer respectively, and the output end of the splicing module is connected to the 1×1 convolution module. The output end of the 1×1 convolution module outputs the feature map of the intermediate fusion layer.

5. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1 or 4, characterized in that, The feature map of the intermediate fusion layer in step 4 is: hfmp i = conv(concat(enfmp′ i , defmp i ))(i = 1, 2, …, n) wherein conv() denotes a 1x1 convolution used to reduce the dimension of the feature map, concat() denotes a channel concatenation operation, defmp i denotes the decoded layer feature map, enfmp′ i denotes the encoded feature map optimized by the feature map self-optimization module.

6. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1, characterized in that, In step 5, the hierarchical weighted fusion module is used to perform hierarchical weighted fusion on the feature map of the intermediate fusion layer, specifically including: Step 5-1: Extract the semantic information of the high-level intermediate fusion layer feature map to form the i-1 layer weight feature map Qfmp i-1 : Qfmp i-1 = res((res(conv(hfmp i ))*(res(sigmoid(conv(hfmp i ))) T )*res(hfmp i-1 )) wherein res() represents a feature map deformation operation, conv() represents a 1 x 1 convolution, hfmp i represents an intermediate fusion layer feature map, defmp i represents a decoding layer feature map, and sigmoid represents an activation function: Step 5-2: the weight feature map Qfmp i-1 is weighted to the intermediate fusion layer feature map hfmp i-1 of the lower layer and fused to obtain the fused hierarchical fusion feature map fhfmp i-1 of the lower layer, and the calculation formula is: fhfmp i-1 = (hfmp i-2 + Qfmp i-2 ) + hfmp i-1 Step 5-3: Hierarchical fusion feature map fhfmp i Obtaining hierarchical independent prediction map pre i : pre i = conv(fhfmp i )(i = 1, 2,..., n) Among them, conv() represents 1×1 convolution.

7. The crack detection method based on feature map self-optimization and multi-layer feature step-by-step weighting fusion according to claim 1, characterized in that, The final prediction graph pre in step 6 is: pre = sigmoid(convp(concat(pre i ))(i = 1, 2, …, n) where sigmoid() denotes an activation function, convp() denotes a 3x3 convolution used to generate a prediction map, concat() denotes a channel concatenation operation, pre i is a layered independent prediction map.

Images