吴江春, 王虎虎, 徐幸莲. 基于机器视觉的鸡胴体断翅快速检测技术[J]. 农业工程学报, 2022, 38(22): 253-261. DOI: 10.11975/j.issn.1002-6819.2022.22.027
    引用本文: 吴江春, 王虎虎, 徐幸莲. 基于机器视觉的鸡胴体断翅快速检测技术[J]. 农业工程学报, 2022, 38(22): 253-261. DOI: 10.11975/j.issn.1002-6819.2022.22.027
    Wu Jiangchun, Wang Huhu, Xu Xinglian. Rapid detection technology for broken-winged broiler carcass based on machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(22): 253-261. DOI: 10.11975/j.issn.1002-6819.2022.22.027
    Citation: Wu Jiangchun, Wang Huhu, Xu Xinglian. Rapid detection technology for broken-winged broiler carcass based on machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(22): 253-261. DOI: 10.11975/j.issn.1002-6819.2022.22.027

    基于机器视觉的鸡胴体断翅快速检测技术

    Rapid detection technology for broken-winged broiler carcass based on machine vision

    • 摘要: 为实现肉鸡屠宰过程中断翅鸡胴体的快速检测,提高生产效率,该研究利用机器视觉系统采集了肉鸡屠宰线上的1 053张肉鸡胴体图,构建了一种快速识别断翅缺陷的方法。通过机器视觉装置采集鸡胴体正视图,经图像预处理后分别提取鸡胴体左右两端到质心的距离及其差值(d1、d2、dc)、两翅最低点高度及其差值(h1、h2、hc)、两翅面积及其比值(S1、S2、Sr)、矩形度(R)和宽长比(rate)共11个特征值,并通过主成分分析降维至8个主成分。建立线性判别模型、二次判别模型、随机森林、支持向量机、BP神经网络和VGG16模型,比较模型的F1分数和总准确率,在所有模型组合中,以VGG16模型的F1分数和总准确率最高,分别为94.35%和93.28%,平均预测速度为10.34张/s。利用VGG16建立的模型有较好的分类效果,可为鸡胴体断翅的快速识别与分类提供技术参考。

       

      Abstract: Abstract: Broken-winged chicken carcasses can be one of the most common defects in broiler slaughter plants. Manual detection cannot fully meet the large-scale production, due to the high labor intensity with the low efficiency and accuracy. Therefore, it is a high demand to rapidly and accurately detect broken wings on chicken carcasses. This study aims to realize the rapid inspection of broken-winged chicken carcasses in the progress of broiler slaughter, in order to improve the production efficiency for the cost-saving slaughter line. 1053 broiler carcass images were collected from a broiler slaughter line using a computer vision system. Rapid identification was then constructed for the broken wing defects. Specifically, the front view of the chicken carcass was obtained in the machine vision system. The preprocessing was then deployed to obtain the chicken carcass images without the background, including the weighted average (graying), two-dimensional median filtering (denoising), and iterative (threshold segmentation). The code was also written in the MATLAB platform. After that, a total of 11 characteristic values were calculated, covering the exact distance starting from the left and right ends of the chicken carcass image to the centroid and the difference (d1, d2, and dc), the heights of the lowest point in the two wings and their difference (h1, h2, and hc), the areas of the two wings and ratio of them (S1, S2, and Sr), squareness (R), and width-length ratio (rate). As such, the eight principal components were achieved in the principal component analysis after the reduction of several dimensions. Separately, the principal components and characteristic values were imported into the specific model of linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), random forest (RF), support vector machine (SVM), and BP neural network. Among them, the input parameter of the VGG16 model was from the RGB maps of the chicken carcass with the removed background. Finally, a comparison was made for the F1-scores and total accuracy of each model. Thus, the highest recall rate of 91.13% was achieved in the RF model with the characteristic values as the input parameters among the shallow learning models. The second higher recognition accuracy and F1-score were 96.28% and 92.58%, respectively in the SVM model with the characteristic values as input parameters. The highest total accuracy of model recognition was achieved in the quadratic discriminant and SVM models with the characteristic values as the input parameters, both up to a proportion of 91.78%. Moreover, the F-score and total accuracy of the VGG16 model were the highest among the total model combinations, with respective rates of 94.35% and 93.28%, respectively. In terms of the prediction time of models, the shortest prediction time was obtained in the SVM model with the principal components as the input parameters. Specifically, the capacity was found to determine 353 sample images in 0.000 9 s, with an average speed of 3.92×105 images per second. By contrast, the longest prediction time was observed in the VGG16 model, where 24.46 s to determine 253 sample images, with an average speed of 10.34 images per second. In conclusion, the VGG16 model can be expected to serve as the best classification of broken wings in chicken carcasses.

       

    /

    返回文章
    返回