Abstract: A machine vision system was used to collect 948 three-dimensional images of chicken carcasses on the broiler slaughter line. This study aimed to develop a rapid method for the identification of primary dermatitis in chicken carcasses. The acquired images were preprocessed and segmented into 128 × 128 pixel pictures with grids. A total of 762 pictures of dermatitic skin and 775 pictures of normal skin were selected. A total of 24 feature values were extracted including third-order color moments, mean and variance of gray-level co-occurrence matrix features, Tamura texture features from the 1 537 pictures and the segmentation threshold and area of dermatitis region. Based on dimensionality reduction by principal component analysis (PCA), linear discriminant analysis model, quadratic discriminant analysis model, support vector machine, random forest, back propagation neural network (BPNN) and GoogLeNet models were established, and their classification performances were compared. Among these models, the GoogLeNet model was the most effective in classifying dermatitic skin samples with an overall accuracy of 90.5% and an average detection speed of 122.65 sheets per second. The prediction accuracy of the model for chicken carcasses with dermatitis was 100%, while that for qualified chicken carcasses was 90%.

Key words: machine vision; primary dermatitis of chicken carcass; machine learning; defect detection