Target detection method in foggy condensation environment

Abstract

The invention discloses a target detection method in a foggy condensation environment, and the method comprises the following steps: S1, shooting a target image, and converting the image into an acceptable format of an image model through preprocessing; s2, image defect judgment and feature extraction; s3, fusing the feature maps of different scales by using a feature fusion algorithm of an SPP module; s4, using a YOLO head algorithm to send the feature map into a multi-layer convolution layer so as to predict the position and the category of a target in the image; and S5, outputting the position and category information of the target detected in the image by the YOLOv5. According to the method, the target image can be accurately and efficiently identified, the position and category information of the target can be determined, the problem of detection precision of a traditional target detection algorithm in a special environment is solved, hoisting of a heavy object by a crane in the special environment does not depend on pure manual detection and sensor detection any more, and accurate and efficient hoisting is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a target detection method in a foggy and condensation environment. Background Art

[0002] In recent years, with the continuous development of artificial intelligence technology, combining neural networks, target detection and other technologies with traditional industries can reduce costs, improve efficiency, and increase the production capacity of various industries. At present, it is difficult to successfully identify hooks in high temperature, foggy, and condensation environments such as steel mills and metallurgy. Therefore, it is of great research significance to use target detection technology to determine the lifting position of the hoisted object. The existing target detection and hook detection technologies mainly have the following two problems: 1. It is difficult to perform accurate detection when the resolution is low or the image is not clear; 2. It is difficult to perform manual detection in special environments. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the existing defects and provide a target detection method in a foggy and condensation environment, which can accurately and efficiently identify the target image and determine the position and category information of the target, solve the problem of detection accuracy of traditional target detection algorithms in special environments, and realize that cranes lifting heavy objects in special environments no longer rely on pure manual detection and sensor detection, ensuring accurate and efficient lifting, and can effectively solve the problems in the background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for detecting a target in a foggy and condensing environment, comprising the following steps: S1: Image input, the servo motor drives the camera mechanism to capture the target image, and the image is preprocessed and converted into a format acceptable to the image model, denoted as x; S2: Image defect judgment and feature extraction, image defect judgment is performed on image x; If image x has no defects, image x is subjected to feature extraction by convolutional neural network, and the features in image x are extracted and recorded as f(x); If image x has defects, mark the defective image x and do not perform detection. Add graphic segmentation, image defect optimization and image stitching to the convolutional neural network, and return to the convolutional neural network for feature extraction. S3: feature fusion, using the feature fusion algorithm of the SPP module to fuse feature maps of different scales, denoted as g(f(x)); S4: Prediction box prediction, using the YOLO head algorithm, the feature map g(f(x)) is fed into multiple convolutional layers to predict the location and category of the target in the image; S5: Output results, YOLOv5 outputs the location and category information of the targets detected in the image.

[0005] Preferably, the training process of the image model in S1 is as follows: a: Input the image of the correctly lifted object into the feature extraction network and output the feature map s1; b: Input the line number image into the feature extraction network and output the feature map s2; c: Classify and regress the feature map, remove redundant prediction boxes, and keep only the prediction box that is closest to the output.

[0006] Preferably, the image model training stage uses the Mosaic data enhancement method to perform data enhancement, and the data set is pre-trained in the YOLOV5 network.

[0007] Preferably, the feature map is classified using adaptive frame calculation and adaptive image scaling to obtain the anchor size of the data set.

[0008] Preferably, when using adaptive frame calculation, the model outputs a corresponding prediction frame based on the initial point frame, calculates the gap between it and the GT frame, and performs a reverse update operation to update the parameters of the entire network; adaptive image scaling adds minimal black edges to the scaled image.

[0009] Preferably, the specific process in S4 is divided into position prediction and category prediction: a: The position prediction is represented by the center coordinates, width and height of the bounding box, denoted as b_{i} = (x, y, w, h); b: Category prediction uses the softmax function to normalize the scores of each category, denoted as c_{i} = (P(c_{1}), P(c_{2}), ..., P(c_{n})).

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: the image is input into the image model, and then the image defects are judged. If the image has no defects, feature extraction is performed. The defective image is subjected to graphic segmentation, image defect optimization and image splicing before feature extraction. After feature fusion, the YOLO head algorithm is used to send the features to the multi-layer convolution layer to predict the position and category of the target in the image, and then the position and category information of the target is output. The target image can be accurately and efficiently identified, and the position and category information of the target can be determined. The problem of detection accuracy of the traditional target detection algorithm under special environments is solved, and the crane lifting heavy objects under special environments no longer relies on pure manual detection and sensor detection, thereby ensuring accurate and efficient work. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0012] The present invention can be explained in detail through the following embodiments. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention. In the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right", etc. indicating directions or positional relationships, they only correspond to the drawings of the present application for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction.

[0013] See also Figure 1 The present invention provides a technical solution: a method for detecting a target in a foggy and condensing environment, comprising the following steps: S1: Image input, the servo motor drives the camera mechanism to capture the target image, and the image is preprocessed and converted into a format acceptable to the image model, denoted as x; S2: Image defect judgment and feature extraction, image defect judgment is performed on image x; If image x has no defects, image x is subjected to feature extraction by convolutional neural network, and the features in image x are extracted and recorded as f(x); If image x has defects, mark the defective image x and do not perform detection. Add graphic segmentation, image defect optimization and image stitching to the convolutional neural network, and return to the convolutional neural network for feature extraction. It should be noted that the convolutional neural network (CNN) architecture is a combination of the CSP (Cross Stage Partial) structure and a lightweight neural network such as Darknet53, which is used to extract features from images; S3: feature fusion, using the feature fusion algorithm of the SPP module to fuse feature maps of different scales, denoted as g(f(x)); S4: Prediction box prediction, using the YOLO head algorithm, the feature map g(f(x)) is fed into multiple convolutional layers to predict the location and category of the target in the image; The specific process is divided into position prediction and category prediction: a: The position prediction is represented by the center coordinates, width and height of the bounding box, denoted as b_{i} = (x, y, w, h); b: Category prediction uses the softmax function to normalize the scores of each category, denoted as c_{i} = (P(c_{1}), P(c_{2}), ..., P(c_{n})) S5: Output results, YOLOv5 outputs the location and category information of the targets detected in the image.

[0014] Through the above steps, the location and category information of the target can be output, the target image can be accurately and efficiently recognized, the location and category information of the target can be determined, and the problem of detection accuracy of traditional target detection algorithms in special environments can be solved. In special environments, cranes no longer need to rely on pure manual detection and sensor detection to lift heavy objects, ensuring accurate and efficient work.

[0015] In this embodiment, the training process of the image model in the above step S1 is as follows: a: Input the image of the correctly lifted object into the feature extraction network and output the feature map s1; b: Input the line number image into the feature extraction network and output the feature map s2; c: Classify and regress the feature map, remove redundant prediction boxes, and keep only the prediction box closest to the output; During the image model training phase, the Mosaic data enhancement method is used for data enhancement, and the dataset is pre-trained in the YOLOV5 network; The feature map is classified using adaptive frame calculation and adaptive image scaling to obtain the anchor size of the data set; when using adaptive frame calculation, the model outputs the corresponding prediction frame based on the initial point frame, calculates the gap between it and the GT frame, and performs a reverse update operation to update the parameters of the entire network; In actual use, many images have different aspect ratios. The sizes of the black borders at both ends after scaling and filling are different, which will result in a large amount of information redundancy and affect the reasoning speed of the entire algorithm. Adaptive imagescaling can add the least black borders to the scaled image.

[0016] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and it is intended that all changes that fall within the meaning and scope of equivalent elements are included in the content of the present invention.

Claims

1. A method for detecting a target in a foggy and condensing environment, characterized in that: The following steps are involved: S1: Image input, the servo motor drives the camera mechanism to capture the target image, and the image is preprocessed and converted into a format acceptable to the image model, denoted as x; S2: Image defect judgment and feature extraction, image defect judgment is performed on image x; If image x has no defects, image x is subjected to feature extraction by convolutional neural network, and the features in image x are extracted and recorded as f(x); If image x has defects, mark the defective image x and do not perform detection. Add graphic segmentation, image defect optimization and image stitching to the convolutional neural network, and return to the convolutional neural network for feature extraction. S3: feature fusion, using the feature fusion algorithm of the SPP module to fuse feature maps of different scales, denoted as g(f(x)); S4: Prediction box prediction, using the YOLO head algorithm, the feature map g(f(x)) is fed into multiple convolutional layers to predict the location and category of the target in the image; S5: Output results, YOLOv5 outputs the location and category information of the targets detected in the image.

2. The target detection method in a foggy and condensing environment according to claim 1, characterized in that: The training process of the image model in S1 is as follows: a: Input the image of the correctly lifted object into the feature extraction network and output the feature map s1; b: Input the line number image into the feature extraction network and output the feature map s2; c: Classify and regress the feature map, remove redundant prediction boxes, and keep only the prediction box that is closest to the output.

3. The target detection method in a foggy and condensing environment according to claim 2, characterized in that: The image model training phase uses the Mosaic data enhancement method to perform data enhancement, and the data set is pre-trained in the YOLOV5 network.

4. The target detection method in a foggy and condensing environment according to claim 2, characterized in that: The feature map is classified using adaptive frame calculation and adaptive image scaling to obtain the anchor size of the dataset.

5. The target detection method in a foggy and condensing environment according to claim 4, characterized in that: When using adaptive frame calculation, the model outputs the corresponding prediction frame based on the initial point frame, calculates the gap between it and the GT frame, and performs a reverse update operation to update the parameters of the entire network; Adaptive image scaling adds minimal black borders to scaled images.

6. The target detection method in a foggy and condensing environment according to claim 1, characterized in that: The specific process in S4 is divided into position prediction and category prediction: a: The position prediction is represented by the center coordinates, width and height of the bounding box, denoted as b_{i} = (x, y, w, h); b: Category prediction uses the softmax function to normalize the scores of each category, denoted as c_{i} = (P(c_{1}), P(c_{2}), ..., P(c_{n})).

Images