Application of Fisheye Image Correction in Intelligent Retail Containers

In recent years, image detection based on deep learning has become one of the main technologies of intelligent retail container (IRC). Fisheye lens is widely adopted as the imaging equipment of the IRC due to its short focal length, large viewing angle and small volume. Aiming at the distortion of fisheye lens imaging, an innovative "center coordinate correcting and clustering algorithm (CCCCA)" based on spherical double longitude model is proposed to correct the classification error of the hard sample in fisheye image predicted by neural network model. First, the YOLOv4Tiny model is used to detect the fisheye image of the IRC to gain the bounding boxes ("bboxes"). Second, the center and radius of the fisheye image are obtained by using the Hough circle algorithm in OpenCV, and the fisheye image is orthogonally mapped to the target plane by means of the spherical double longitude model so as to get the correction center coordinates of the bboxes. Finally, the x-axis coordinates of the correction bbox’s centers are clustered to determine the categories of the hard samples ("HardSampleSet"). Experimental results show that the CCCCA can reduce the top-1 error rate ("top-1 err.") of the HardSampleSet in our project by 5.57%.