A Generative Adversarial Network-based Framework for Fruit and Vegetable Occlusion Detection in Smart Refrigerators

With the development of deep learning, image recognition technology has made great progress. However, there is often occlusion in the image recognition task. Object occlusion not only loses part of the target information, but also introduces additional interference, thus exacerbating the difficulty of image recognition. This paper aims to improve the recognition rate of fruits and vegetables in the presence of occlusion, so as to alert people to the timely disposal of food in the refrigerator when it is nearing its expiration date. To this end, this paper employs the Alexnet architecture and revises it for better feature extraction, and combines it with a generative adversarial network (GAN), which trains a generator and a discriminator with pairs of occluded and non-occluded images, and finally recover the occluded images. Experimental results show that the proposed system improves the accuracy of fruit and vegetable recognition, and can be better used in smart refrigerators to remind the shelf life of fruits and vegetables.