computer vision detection of salmon muscle gaping using convolutional neural network features

Salmon muscle gaping will lead to the irregular voids or undesirable lace-like appearance in the final product. this study was carried out to develop an automatic imaging analysis method for rapid, accurate, and non-invasive detection of gaping blemishes on salmon carcasses. salmon fillets could be classified as wholesome or defective samples based on the number of candidate gaping regions in the preliminary step applying local adaptive thresholding. supervised classification results were compared between using histograms of oriented gradients (hog) and convolutional neural network (cnn) feature extractors. it was shown that cnn features outperformed hog features with correct classification rates (ccrs) of 0.927 and 0.916 for cross validation and test data set, respectively. relieff was then applied to select important feature attributes by reducing the 4096-dimensional to 239-dimensional vector. simplified cnn model also yielded good classification performance with ccr of 0.925 for cross validation. therefore, cnns were used to extract features from candidate regions and then reduced features to the 239-dimensional vector. the resultant vector was fed to the simplified cnn model to make a final decision. the prediction maps for visualizing the classification result on salmon fillet were subsequently generated. the overall results confirmed that this proposed method is effective and suitable for the muscle gaping detection. future work will be focused on applying this method in packing plants where fish fillets are progressing rapidly, and promising results will allow the identification of critical points in the supply chain that impact upon product quality.