Improving Fish Freshness Classification with Fish Eye Segmentation Guidance

Estimating fish freshness plays an essential role in multiple fields, including the food industry. Most existing freshness estimation methods progress in a biochemical manner, which can inevitably cause defects in the fish samples. Although recently advanced image processing-based approaches realize this task from fish images, achieving satisfactory accuracy is still difficult because fish bodies do not often reflect freshness well. In this study, inspired by the fact that fish eyes possess a strong correlation with freshness, we propose a simple yet effective method that learns to classify fish freshness into three predetermined levels from fish eye images. Specifically, we first train a segmentation network to produce a mask that identifies the fish eye area. This mask is then concatenated to the fish eye image as an additional input channel for the classification network to output the freshness level. Consequently, the classification network is explicitly guided to focus more on the fish eye area than on the background. Experimental results on a publicly available fish eye dataset demonstrate that our method contributes to improved freshness classification accuracy in terms of the four types of classification backbones.