Deep Learning Based Supermarket Product Detection and Recognition with Rectified Images

Object Detection has become a very competent domain of computer vision off late. Many new models have paved way for the research to expedite exponentially. Complex models like YOLO, RetinaNet, RCNN has revolutionised the domain completely. Although the trend of creating deeper models with huge computational complexity is on the rise, the traditional computer vision algorithms still hold their own. One such traditional concept of calculating image homography has passed the test of time. In this paper we study the effect of rectifying images using image homography on the problem of grocery identification. One major problem in implementing vision algorithms for grocery identification is that the cameras installed in the stores focuses more on the pavements than on the products, thus generating angled images of the products in shelf. The same problem lies when we try and capture the images using mobile phones. Thus in order to resolve this problem of side facing of images we use homography estimation coupled with hough transform to give us frontal alignment of the images. We captured 1500 images containing a total of 590 classes and trained model with these images. For testing we used 200 self captured angled images (100 right aligned and 100 left aligned) of different products. The results obtained showcases that the rectification of images show a marginal improvement in accuracy than the non rectified images.