CN-116524232-B - Improved YOLOv-Tiny realization engineering machinery and material identification method

CN116524232BCN 116524232 BCN116524232 BCN 116524232BCN-116524232-B

Abstract

The invention relates to the field of artificial intelligence and discloses an improved YOLOv-Tiny recognition method for realizing engineering machinery and materials, which comprises the steps of collecting engineering machinery and material pictures, establishing a dataset sample, wherein the dataset sample is used for training an improved YOLOv-Tiny model, establishing an improved YOLOv-Tiny model, finally adding an SPP multi-scale feature fusion module into a trunk feature extraction network of an original YOLOv-Tiny network to realize the expansion of a receptive field, inserting an SE channel attention module in front of the SPP multi-scale feature fusion module to realize the reduction of irrelevant information interference, adjusting an FPN feature pyramid fusion module, enhancing the utilization of shallow information to realize the enhancement of network characterization capability, and using the dataset sample for model training based on the improved YOLOv-Tiny model to realize the detection and classification of engineering machinery and material targets. Experimental results show that the algorithm of the invention can enable the detection and classification of engineering machinery and materials to be more accurate.

Inventors

AI CHUN
WANG XUDONG
LIU SHUYING
GAO GEN

Assignees

缤谷电力科技（上海）有限公司

Dates

Publication Date: 20260508
Application Date: 20230320

Claims (9)

1. An improved YOLOv-Tiny method for realizing engineering machinery and material identification is characterized by comprising the following steps: collecting engineering machinery and material pictures, and establishing a data set sample, wherein the data set sample is used for training an improved YOLOv-Tiny model; Establishing the improved YOLOv-Tiny model, adding an SPP multi-scale feature fusion module into a main feature extraction network of an original YOLOv-Tiny network to enlarge a receptive field, inserting an SE channel attention module in front of the SPP multi-scale feature fusion module to reduce irrelevant information interference, adjusting an FPN feature pyramid fusion module to strengthen the utilization of shallow information so as to enhance network characterization capability, adjusting the connection mode of an original feature fusion structure in the improved FPN feature pyramid fusion module, adding a plurality of fusion channels on the basis of the original network, wherein a feature map output by the SPP multi-scale feature fusion module enters a first channel and a second channel after passing through two convolution blocks, and enters a third channel and a fourth channel after passing through a CSPBlock residual module for downsampling to obtain a feature map; After the feature map transmitted by the channel three passes through a convolution block with the step length of 2, the feature map is spliced with the feature map transmitted by the channel one for the first time, then the channel fusion is carried out through a 3*3 convolution block, and finally the feature map is spliced with the feature map transmitted by the channel two for the second time, and then the feature map is output; A channel five is arranged behind the 3*3 convolution blocks, the feature images after the first splicing are transmitted through the channel five, the feature images transmitted by the channel five are spliced with the feature images transmitted by the channel four after passing through one convolution block and performing up-sampling operation to enlarge the size, and finally a new feature image is output; Based on the improved YOLOv-Tiny model, model training is carried out by using the dataset sample so as to realize detection and classification of engineering machinery and material targets.
2. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 1, wherein the improved YOLOv-Tiny model comprises a trunk feature extraction network and a feature fusion part, wherein the trunk feature extraction network is used for realizing extraction of input image features, the feature fusion part is used for realizing multi-scale feature fusion of a feature map, the trunk feature extraction network comprises a CBL convolution block and a CSPBlock residual module, and the feature fusion part comprises an SPP multi-scale feature fusion module, an SE channel attention module and the FPN feature pyramid fusion module.
3. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 2, wherein the CBL convolution block comprises a Conv convolution layer, a BN batch normalization layer and a Leaky Relu activation function, the Conv convolution layer is arranged before the BN batch normalization layer, the Leaky Relu activation function is arranged after the BN batch normalization layer, wherein the Conv convolution layer is used for extracting local space information in input features to obtain different response feature diagrams, the BN batch normalization layer is used for maintaining relatively stable distribution of input data of each layer in a network and accelerating model learning speed, and the Leaky Relu activation function is used for solving the problem that neurons are not learned after the Relu function enters a negative interval.
4. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 2, wherein the CSPBlock residual module divides input features into two parts, wherein one part of the input features is introduced into a Route fusion layer to be spliced with the other part of the input features after passing through a residual block formed by 3*3 convolution, so that the improved YOLOv-Tiny model learns more features.
5. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 4, wherein the CSPBlock residual error module is followed by a Maxpool maximum pooling layer, and the Maxpool maximum pooling layer is used for downsampling, removing redundant information, compressing features and simplifying network complexity.
6. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 2, wherein the SE channel attention module comprises a Avgpool average pooling layer, two FC full-connection layers and a Sigmoid layer, wherein the Avgpool average pooling layer is used for preserving the characteristics of overall data and projecting background information, the two FC full-connection layers are used for learning channel characteristics of the compressed characteristic map so as to obtain the characteristic map with channel attention, and the Sigmoid layer is connected after the second FC full-connection layer and is used for reducing divergence of data in the transmission process.
7. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as claimed in claim 6, wherein the number of neurons in the first FC full-connection layer is C/r, wherein C is the number of neurons and r is a scaling factor, r=16 is adopted for dimension reduction, and the second FC full-connection layer is used for dimension lifting.
8. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny implementation method according to claim 5, wherein the SPP multi-scale feature fusion module comprises a jump connection and three different convolution kernel sizes of Maxpool maximum pooling layers, the jump connection is used for outputting an uncolored operation feature map, the three different convolution kernel sizes of Maxpool maximum pooling layers output three different pooled results, and the Route fusion layer and the uncolored operation feature map are fused to output a multi-scale feature map, and the fused multi-scale feature map is used for adjusting the channel number by using 1*1 convolution operations and enters the improved FPN feature pyramid fusion module.
9. The method for realizing engineering machinery and material identification by using the improved YOLOv-Tiny as set forth in claim 1, wherein the data set samples are divided into a training set, a verification set and a test set, the training set is used for training the improved YOLOv-Tiny network model, the verification set is used for adjusting parameters of the improved YOLOv-Tiny network model, and the test set is used for testing whether the improved YOLOv4-Tiny network model is accurate or not.

Description

Improved YOLOv-Tiny realization engineering machinery and material identification method Technical Field The invention relates to the field of artificial intelligence, in particular to an improved YOLOv-Tiny recognition method for realizing engineering machinery and materials. Background Along with the development of science and technology and the maturation of artificial intelligence technology, various industries are developing towards the trend of intelligence, wherein the intelligent development of engineering machinery is attracting a great deal of attention, and becomes one of research hotspots. The engineering machinery is taken as one of important prop industries in China, plays a vital role in the development of infrastructure construction in China, covers the fields of infrastructure, mining and the like, is severe in working environment, often accompanies high temperature, dust, vibration and the like, is a serious threat to the safety of workers caused by complex working environment, and has higher repeatability and a fixed working area in part of work. Therefore, in order to reduce the safety risk of engineering machinery operation on people, the labor cost is saved, the development of unmanned and intelligent engineering machinery has become a trend, the working risk can be reduced, the working efficiency of a loader is greatly improved, and the method has important practical significance, but the detection and classification of the engineering machinery and materials by the existing artificial intelligence technology are easily interfered by irrelevant information. Disclosure of Invention The invention aims at solving the technical problems and provides an improved YOLOv-Tiny recognition method for realizing engineering machinery and materials, which can be realized by the following technical scheme: an improved YOLOv-Tiny method for realizing engineering machinery and material identification is characterized by comprising the following steps: collecting engineering machinery and material pictures, and establishing a data set sample, wherein the data set sample is used for training an improved YOLOv-Tiny model; Establishing the improved YOLOv-Tiny model, adding an SPP multi-scale feature fusion module into a main feature extraction network of an original YOLOv-Tiny network to expand a receptive field, inserting an SE channel attention module in front of the SPP multi-scale feature fusion module to reduce irrelevant information interference, adjusting an FPN feature pyramid fusion module, and enhancing the utilization of shallow information to enhance the network characterization capability; Based on the improved YOLOv-Tiny model, model training is carried out by using the dataset sample so as to realize detection and classification of engineering machinery and material targets. Further, the improved YOLOv-Tiny model comprises a trunk feature extraction network and a feature fusion part, wherein the trunk feature extraction network is used for extracting input image features, the feature fusion part is used for realizing multi-scale feature fusion of a feature map, the trunk feature extraction network comprises a CBL convolution block and a CSPBlock residual module, and the feature fusion part comprises an SPP multi-scale feature fusion module, an SE channel attention module and an FPN feature pyramid fusion module. Further, the Conv convolution layer is arranged before the BN batch normalization layer, the Leaky Relu activation function is arranged after the BN batch normalization layer, and the Leaky Relu activation function is arranged after the BN batch normalization layer for use, wherein the convolution layer is used for extracting local spatial information in input features to obtain different response feature graphs, the BN batch normalization layer is used for maintaining relatively stable distribution of input data of each layer in a network and accelerating model learning speed, and the Leaky Relu activation function is used for solving the problem that neurons are not learned after Relu functions enter a negative interval. Further, the CSPBlock residual module divides the input feature into two parts, wherein one part of the input feature is introduced into a large residual edge while passing through a residual block formed by 3*3 convolution, and then enters a Route fusion layer to be spliced with the other part of the input feature, so that the improved YOLOv-Tiny model learns more features. Further, the CSPBlock residual module is followed by a Maxpool max pooling layer, and the Maxpool max pooling layer is used for downsampling, removing redundant information, compressing features, and simplifying network complexity. Further, the SE channel attention module includes Avgpool an average pooling layer, two FC full-connection layers and a Sigmoid layer, where the Avgpool average pooling layer is used to retain the characteristics of overall data and project background information,