Abstract:
Abstract: Poultry carcass defects and condemnation (such as bruises) have posed a great threat to the broiler industry in recent years. Among abnormal quality, the broiler carcass congestion has also brought serious economic losses to slaughtering enterprises. It is an urgent need to establish a rapid and accurate identification of carcass congestion in the industry. In this study, an image acquisition device was developed for the rapid detection of broiler carcass congestion using machine vision. A three-directional visual acquisition system (equipped with three light sources) was adopted to realize the full coverage of the broiler carcass in the field of view. The image was also captured and then preprocessed using the grayscale, Gaussian filter denoising, binarization, and morphological processing. The maximum circumscribed rectangle of the carcass was obtained, and then the image was divided into the front and side view. Firstly, the global RGB color threshold was determined to segment the congestion in the carcass image. 14 characteristic parameters of the image were also extracted, including the R, G, B, H, I, S, a*, b*, Ar1 (the total area of the first type of congestion), Ar2 (the total area of the second type of congestion), Pc1 (the percentage of the first type of congestion), Pc2 (the percentage of the second type of congestion), M1 (the largest area in the first type of congestion), and M2 (the largest area in the second type of congestion). A Pearson correlation analysis was performed on the characteristic parameters. A principal component analysis was implemented to obtain the seven principal components after dimension reduction. The classification model was trained using a Support Vector Machine (SVM) combined with the Genetic Algorithm (GA). Secondly, the maximum circumscribed rectangular images of carcasses were traversed using the sliding window. The sub-images of (50 × 50) pixels were then divided into the images. The calibration coefficient was integrated to determine the real area of a sub-image (6 cm2) for each image. The sub-images were then divided into four categories, namely congestion, normal skin, carcass-background junction, and background. Four types of large sub-images with outstanding characteristics were manually selected to extract the color moment. The SVM-GA model was also achieved in this case. Finally, the similarity measure was used to revise the classification of the model. The results show that the classification accuracies of the SVM model using seven principal components were 86.0% and 89.8%, respectively, in the front and side view. The prediction time was 0.006 s. Nevertheless, there was no ideal effect of RGB threshold segmentation. By contrast, the classification accuracies of the SVM model using color moment were 98.3% and 97.9%, respectively, where the prediction time was 0.001 s. Furthermore, the recognition recall rate of congestion in the test sample was improved significantly after the revision of the model combining the Euclidean distance with the similarity measure. More importantly, there was a great decrease in the false positive rate and the miss rate. Consequently, the trained SVM model with the similarity measurement can be expected to effectively identify the carcass congestion using the local color moment of carcass sub-images, compared with the global RGB threshold segmentation at present. The finding can provide a strong reference to conduct the real-time detection of carcass congestion in poultry production.