Abstract
To solve the problems existing in the traditional algorithm, such as image quality, the ability to predict the coordinate of the motion fat reduction image, the Intersection over Union (IOU) of the estimated coordinate of the landmark, the peak signal-to-noise ratio (PSNR) of the image, and the recognition accuracy of the image landmark, an automatic recognition algorithm for fat reduction motion image landmark using computer vision is proposed. The OV7725 camera is used to obtain the fat reduction motion image, and the histogram equalization method is used to enhance the fat reduction motion image. After that, the estimated coordinates of the landmark are obtained by the automatic recognition model of the landmarks of the fat reduction motion image. According to the coordinates of the landmark, after using a field programmable gate array to segment the image of the fat reduction motion image, the incremental updated freeman chain code extracts the contour of the landmark of the fat reduction motion image through 8-neighborhood retrieval, so as to realize the automatic recognition of landmark. The results show that the fat reduction motion image enhanced by the algorithm is clear and the contrast is obvious. The ability to predict the coordinates of the landmark of the fat reduction image is strong. The IOU of the estimated landmark coordinates is about 0.96. The processed image has a high PSNR, which can effectively recognize the landmark of the fat reduction motion image, and the recognition accuracy is about 97%, which has certain value in the field of image recognition and can improve the fat reduction movement normality.
1. Introduction
Automatic recognition of image landmarks is an important image processing technology in the fields of photogrammetry, optical measurement, and digital image processing [1, 2]. With the continuous development of video technology, the requirements for the recognition accuracy of image landmark are more strict in the process of image processing [3]. In order to improve the automatic recognition effect of landmarks in image processing [4], many experts at home and abroad have proposed landmarks recognition methods. Chen et al. [5] proposed an image landmark recognition algorithm based on residual multitask learning. A residual learning module is designed to further strengthen the connection between the two learning tasks as a way to build a residual multitask learning architecture. A training dataset with multitask labels is constructed, the dataset with single-task labels can be used for multitask learning tests, and the residual multitask learning after the tests is used for image landmark recognition. However, the method does not preprocess the images such as denoising, resulting in less accurate final results. Geraa and Balasubramanian [6] proposed an end-to-end recognition architecture for face recognition algorithms based on contextual information, which obtains local and global attention for each channel at each spatial location through a novel spatial channel attention network without obtaining any information from landmark detectors. Scanning is complemented by the complementary contextual information branch, which constructs expression representations from pretrained face recognition features. Redundancy in face recognition features is eliminated by using effective channel attention. However, this method can only recognize facial features and is not accurate for image landmark recognition pairs. Mousavi et al. [7] proposed a landmark-based face recognition algorithm. The new features are proposed based on the number of key points and the most significant landmark regions of the face obtained by the improved scale-invariant feature transformation algorithm. Improved genetic algorithms are used to optimize the weights of features. A support vector machine classifier is used to classify and recognize the weighted features. However, this method is only able to recognize facial features, and the accuracy is not high for image landmark recognition pairs. Fard et al. [8] proposed a knowledge extraction-based neural network algorithm for face landmark recognition as a new loss function that is used to train lightweight student networks (e.g., MobileNetV2) for facial landmark detection. Facial landmarks predicted by two pretrained teacher networks are used to guide the lightweight student network to predict hard landmarks as a way to identify face landmarks. However, this method requires manual photographing, which is wasteful of manpower and not suitable for massive image landmark recognition. Pan et al. [9] utilize graph matching and pseudo-labeling for joint face detection and face landmark recognition. In this paper, a real-time framework for joint face detection and face landmark localization is proposed. Using the synergy between the two tasks of face detection and face landmark localization, a fully convolutional network is designed to predict the locations of face landmarks and face regions. However, since the method does not perform any processing on the image, it leads to a decrease in the subsequent recognition accuracy.
Computer vision means to use cameras and computers instead of human eyes to recognize and measure images [10, 11]. At present, computer vision is used in many fields such as traffic navigation, target tracking, target recognition, and unmanned driving. Since the images acquired by the traditional algorithm have certain noise, which leads to the decrease of subsequent recognition accuracy, this paper proposes a new automatic recognition algorithm for fat reduction motion images. The main contributions of this paper are as follows: (1) The fat reduction motion images are obtained based on computer vision technology that can improve the speed and quality of image acquisition and lay a solid foundation for the subsequent automatic recognition of fat reduction motion image landmarks. (2) After using the histogram equalization method to enhance the fat reduction motion image, the estimated coordinates of the landmark can be obtained through the automatic recognition model of the fat reduction motion image landmark, which can improve the accuracy of the estimated coordinates. (3) After segmenting the fat reduction motion image using the field programmable gate array (FPGA) according to the coordinates of the landmarks, the incremental updated Freeman chain code extracts the contours of the landmarks of the fat reduction motion image through the 8-neighbor search method to realize the automatic recognition of the landmarks and the 8-neighbor search method. The experimental results of the proposed algorithm illustrate the effectiveness of the algorithm in this paper.
2. Methodology
2.1. Fat Reduction Motion Image Acquisition
The acquisition of fat reduction motion image is the basis of its automatic recognition of landmarks. Here, the OV7725 camera is used to capture the fat reduction motion image. The camera has the ability of color image acquisition, and its photosensitive array is 640∗480, which has high sensitivity and is more suitable for low illumination shooting environment. The OV7725 camera has automatic adjustment function, which can effectively suppress the interference of noise points in the image [12, 13].
2.2. Low-Frequency Fat Reduction Motion Image Enhancement
When the OV7725 camera takes the motion image of fat reduction personnel, it is affected by the shooting environment and fat reduction motion behavior, resulting in a fuzzy edge of the fat reduction motion image and an uneven overall gray level of the image [14]. The histogram equalization method is used to enhance the low-frequency image of fat reduction motion. The process is as follows.
Let be the gray level of the fat reduction motion image, and the gray level ranking of the fat reduction motion image is represented by . is the probability density function of the fat reduction motion image, and its expression equation is as follows:where represents the gray level serial number and its value range is 0–255. represents the number of pixels in the gray level. is the sum of pixels of the whole fat reduction motion image.
Let be the setting threshold, and substituting the threshold into equation (1), we have
When , the value remains unchanged. When , the value is the same as the set threshold.
Taking the result of equation (2) as the constraint condition, the fuzzy fat reduction motion image is enhanced by histogram equalization algorithm. Let be the gray level sequence number [15] after enhancement of the fuzzy fat reduction motion image, and its equation is as follows:where represents the rounding operator and and represent the fuzzy fat reduction motion image, respectively.
According to the previous equation, the selected threshold has a great influence on the enhancement effect of fuzzy fat reduction motion image [16, 17]. Here, the fuzzy statistical theory is used to select the appropriate threshold. Let be the gray value of the pixel element of the fat reduction motion image, and the number of blur is represented by . Using the fuzzy membership function to fuzzify the probability density of the motion image, is the result of the fuzzification process. We havewherewhere represents the regularization factor.
Calculating the first-order partial derivative of the optimized histogram of the probability density of the fat reduction motion image [18], the expression equation is as follows:where represents variable parameters.
Let be the local maximum gray level sequence number of fat reduction motion image, and its expression equation is as follows:
After using the histogram of fat reduction motion image optimized by the median filter, the appropriate threshold value is obtained and its equation is as follows:where represents the number of local maximum values and .
Using the results of equations (8)–(10) to enhance the fuzzy fat reduction motion image and obtain the enhanced fat reduction motion image, the expression equation is as follows:where and represent the enhanced fat reduction motion image and histogram equalization value, respectively.
After the above steps, the fuzzy low-frequency fat reduction motion image can be enhanced to make the image clearer and the contrast more obvious, so as to provide a good image basis for the follow-up movement trajectory of its landmarks.
2.3. Construction of an Automatic Recognition Model of Fat Reduction Motion Image Landmarks
The landmarks are given in the fat reduction motion image after image enhancement, which is represented by , and the coordinate of the landmarks in the fat reduction motion image is unknown. Since there is a certain time sequence law in the shooting process of a fat reduction motion image [19], the estimated coordinates of landmarks in its image are calculated by using the interpolation of known coordinate points in the fat reduction motion image. The expression equation of the interpolation polynomial model of medium is as follows:where represents the estimated coordinates of the landmarks in its image. represents the interpolation polynomial model.
To avoid the large deviation between the estimated coordinates and the actual coordinates of the marking edge of the fat reduction motion image [20], the results of equation (11) need to be adjusted according to the motion law of the image marking point. When there are jumping and other actions in the fat reduction motion image, equation (1) is replaced by a first-order model; that is, the value of is 1. and represent the error values of row and column coordinates respectively; then, the error calculation equation of the model is as follows:where and represent the estimated row and actual row of the landmarks in the -th image, respectively. and represent the estimated column and actual column of the flag point in the -th image, respectively.
According to the estimated coordinates of the fat reduction motion image as the center point [21], we draw the rectangular search range of landmarks.
Let and be the width and height of the rectangular search range, respectively, and the expression equation is as follows:where and represent the height and width of the maximum landmarks in the fat reduction motion image, respectively. and represent the maximum error values of landmark rows and columns in the previous image, respectively. is the motion estimate.
According to the result of equation (11), the landmark area of the fat reduction motion image is segmented from the oblique upper coordinate point of the search area as the image segmentation point.
Let and denote the coordinates of the rows and columns corresponding to the starting point of the rectangular search in the first fat reduction motion image [22], respectively, and its expression equation is as follows:where the equation holds when and is greater than and respectively. Otherwise, the values of and are 0.
When recognizing the landmark of the reduced fat motion image through computer vision, after searching for the starting point of the area according to the landmark obtained by equation (14), the reduced fat motion image is input into the FPGA installed in the computer. When recognizing the landmark of the fat reduction motion image, the FPGA calls the fat reduction motion image from the memory card and uses the parallel mode to segment it. The segmentation steps are as follows: Step 1: set the estimated row change interval of the landmarks, and call the estimated coordinate value of the landmarks of the fat reduction motion image in the change interval Step 2: retrieve the fat reduction motion image in the memory card, and temporarily store the image in the temporary storage module of the programmable gate array Step 3: use the fat reduction motion image obtained in the first step to estimate the coordinate value, and segment the search range from the fat reduction motion image in order
After the fat reduction motion image is segmented according to the above steps, it is stored in the double port BRAM storage mode.
The incremental update Freeman chain code method is used to realize the landmark recognition of fat reduction motion image. The incremental updated Freeman chain code extracts the landmark contour of fat reduction motion image by 8-neighborhood retrieval. The process is as follows: Step 1: take the landmark of a fat reduction motion image as the center, traverse in 8 directions, obtain the position information of contour coincidence points, connect the coincident contour points, take any landmark other than the center point as the center, repeat the above operation, and connect the coincident contour points. Place the coincident contour point at the initial position of the connection point group. Step 2: traverse the noncoincident contour points. The noncoincident contour point array and the coincident contour point array are combined to obtain a complete set of connection points to realize the recognition of landmark in the fat reduction motion image, and the recognition results are transmitted to the user’s PC and presented to the user.
Based on the above steps, the automatic recognition process of fat reduction motion image landmarks is shown in Figure 1.
3. Experimental Analysis and Results
3.1. Experimental Environment and Dataset
The experiments used in this paper are carried out on a computer with 10 core Intel Xeon E5-2640 CPU, 64 GB memory, and Windows 10 operating system. The hard disk used in the experiment is HDD 10TB and SSD 480 GB, and the simulation software is MATLAB7.2.
This experiment includes two public datasets and one nonpublic dataset. Public datasets include Pascal VOC dataset and MS COCO dataset. The Pascal VOC dataset is the ancestor of visual recognition competition, including object classification, target detection, image segmentation, and other tasks. The full name of MS COCO dataset is Microsoft Common Objects in Context, which originated from the Microsoft COCO dataset funded and labeled by Microsoft in 2014. The nonpublic dataset takes a college student as the experimental object and uses the OV7725 camera to collect the students’ fat reduction motion images as the nonpublic dataset.
The data in the three datasets are integrated, and the images are divided into two datasets, marked as dataset A and dataset B, respectively. The fat reduction motion images in dataset A are taken under strong light conditions, and the fat reduction motion images in dataset B are taken under low light conditions. The number of student fat reduction motion images in the two datasets is the same, which is five thousand sample images. The data categories include yoga, hiking, jogging, cycling, swimming, rope skipping, and aerobics. Five hundred sample images are extracted from dataset A and dataset B, respectively. 70% of the sample images are used as training data, and 30% of the images are used as test data. The proposed algorithm is used to recognize the landmarks in the images in dataset A and dataset B.
3.2. Evaluation Index
The automatic recognition algorithm of a circular landmark in literature [5], the water gauge mark recognition algorithm in literature [6], the image mark feature adaptive recognition algorithm in literature [7], the mark automatic recognition algorithm in literature [8], the motion image sequence target recognition algorithm in literature [9], and the proposed algorithm are compared, and the practical application effects of different algorithms are verified through different experimental indexes:(1)Image contrast: in order to better present the image enhancement ability of this algorithm, an image from dataset B is selected as the experimental object and the contrast is used as the measurement index to further test the image enhancement ability of this algorithm.(2)Difference in the number of pixels between estimated coordinates and actual coordinates: the lower the difference between the number of pixels in the estimated coordinates and the actual coordinates, the stronger the ability to predict the coordinates of the landmarks in the motion fat reduction image.(3)IOU: IOU refers to the overlap between the estimated landmark coordinates and the actual landmark coordinates. It is one of the indicators to measure the accuracy of the estimated landmarks. The higher the value, the more accurate the estimated landmark coordinates of the fat reduction motion image.(4)PSNR: PSNR is an objective standard for evaluating images. The calculation equation of this index is as follows: Here, is the mean square error. represents the maximum value of the image point color.(5)Recognition effect of fat reduction motion image landmarks: the smaller the number of unrecognized landmarks, the better the recognition effect of landmarks in fat reduction motion images.
3.3. Results and Discussion
The image contrast test results are shown in Figure 2.
(a)
(b)
Analysis of Figure 2 shows that the initial motion fat reduction image does not have enough light and darkness and the image is fuzzy, but after the processing of the proposed algorithm, the overall brightness and darkness of the drawing is increased and the contrast between the black area and the white area is greatly improved, making the image more clear and full.
The estimated coordinates of the landmarks are the basis for subsequent landmark recognition. The pixel points are used as the measurement index, and the pixel difference threshold between the estimated landmark and the actual landmark is set to 10. The threshold value of 10 is because during the testing process using the test set, a pixel difference between the predicted landmarks and the actual landmarks greater than 10 results in a large marker point prediction error. This leads to a decrease in the accuracy of subsequent fat reduction motion image landmark recognition. The proposed algorithm estimates the average pixel difference between the X-axis and Y-axis coordinates of the landmarks when the number of landmarks is different, and the result is shown in Figure 3.
According to Figure 3, when the number of landmarks is different, the average difference between X-axis and Y-axis coordinates estimated by the proposed algorithm is irregular. However, the pixel difference changes little. When there are 60 and 120 landmarks, the maximum estimated pixel difference between X-axis and Y-axis is 8. The value is below the set threshold. In summary, the proposed algorithm has a good ability to predict the coordinates of the logo points of the sports fat reduction image, which also proves that the proposed algorithm has a strong ability to recognize the logo points from the side.
Taking the IOU as a measure of the accuracy index of the landmarks estimated by the proposed algorithm, the change of the IOU of the landmark coordinates estimated by this method under the condition of different number of landmarks is tested and the results are shown in Figure 4.
It can be seen from Figure 4 that with the increase in the number of landmarks, the value of the intersection and merger ratio of the proposed algorithm shows a gentle fluctuation trend. However, its fluctuation range is small, and the value of the IOU is always maintained at about 0.96. The above data show that the proposed algorithm is more accurate in predicting the coordinates of landmarks in fat reduction motion images.
A group of images from dataset A and dataset B are selected, respectively, and marked as image A and image B. We take the PSNR as the measurement index to test the PSNR when the image is enhanced by the proposed algorithm. The test results are shown in Table 1.
According to the data in Table 1, the initial PSNR of fat reduction sports images is different in different shooting environments. However, the initial PSNR values of fat reduction sports images taken under low light and strong light conditions are not much different. After the two images are enhanced by the proposed algorithm, the PSNR values are 19.26 and 15.64, respectively, which are 9.95 and 7.4 higher than the initial PSNR, respectively. In the comparison algorithm, the algorithm in literature [9] has a slightly higher PSNR value after the fat reduction motion image is enhanced than other algorithms. Especially after image B is enhanced, its PSNR value is only the same as that of the proposed algorithm. The difference is 0.61, and its image enhancement capability is second only to the proposed algorithm. In summary, the proposed algorithm has good image enhancement capabilities and can provide a good image basis for the automatic recognition of fat reduction motion image landmarks.
Using the number of recognition landmarks as an index, the effect of the proposed algorithm on recognizing the landmarks of fat reduction motion images is tested. At the same time, it is compared with literature [5] algorithm, literature [6] algorithm, literature [7] algorithm, literature [8] algorithm, and literature [9] algorithm. The result is shown in Figure 5.
(a)
(b)
(c)
(d)
(e)
(f)
According to Figure 5, none of the six algorithms recognized the chin position landmark of the person in the fat reduction motion image. However, the number of landmarks of the fat reduction motion image is 1, which is not recognized by the proposed algorithm, the number of landmarks of the fat reduction motion image is 5, which is not recognized in literature [5], the number of landmarks of the fat reduction motion image is 3, which is not recognized in literature [6], the number of landmarks of the fat reduction motion image is 4, which is not recognized in literature [7], the number of landmarks of the fat reduction motion image is 2, which is not recognized in literature [8], and the number of landmarks of the fat reduction motion image is 4, which is not recognized in literature [9]. Therefore, the proposed algorithm can effectively recognize the landmark of the reduced fat motion image, and the recognition accuracy is 97%.
4. Conclusions
In this paper, computer vision technology is applied to the recognition process of fat reduction motion image landmarks. After the image is enhanced and processed, an automatic recognition model of fat reduction motion image landmarks is constructed and the model is used to realize the automatic recognition of fat reduction motion image landmarks. The results show that the enhanced fat reduction motion image of the algorithm is clear and has high contrast, the ability to predict the coordinates of the marker points of the motion fat reduction image is strong, and the value of the intersection and ratio of the coordinates of the predicted landmarks is as high as about 0.96. The PSNR of the processed image is high, and the recognition accuracy is about 97%, which has achieved a good application effect. However, the convergence of the algorithm used is not verified in this paper. Therefore, the convergence of the algorithm used in this paper needs to be verified in the future research to improve the authenticity and reliability of the research of this paper.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Conflicts of Interest
The authors declare that there are no conflicts of interest.
Acknowledgments
This work was supported by the Heilongjiang University Scientific Research Project Foundation of China under Grant No. 2021KYYWF-PY02.