Detection of Atlantic salmon bone residues using machine vision technology

The detection of foreign bodies in the food industry has received considerable research attention in recent years. This study aimed to assess the efficacy of combining machine vision with neural network models for detecting residual bones in Atlantic salmon flesh as well as explore the degree to which image quality affects the performance of object detection models. We first employed region segmentation and various forms of data expansion to obtain 3120 images of Atlantic salmon bone residues and then used the image compression algorithm to obtain data sets with images that differ in quality. Three object detection models (Faster-RCNN + Alexnet, Faster-RCNN + VGG16 and Faster-RCNN + VGG19) were trained based on uncompressed image data sets. The Faster-RCNN + VGG16 model had optimal test performance on the image data set with a compression ratio of 25% (F1-score = 0.87, AP = 0.78). The Faster-RCNN + VGG16 model for images with a compression ratio of 10% did not show any practical value (F1-score = 0.17, AP = 0.04). Therefore, neural network models based on machine vision can robustly detect residual bones in Atlantic salmon flesh from images containing bones. The quality of images used for detection had a significant impact on the detection results, and small target images should be less robust to compression.