Improving Small Object Proposals for Company Logo Detection

Abstract

Many modern approaches for object detection are two-staged pipelines. The first stage identifies regions of interest which are then classified in the second stage. Faster R-CNN is such an approach for object detection which combines both stages into a single pipeline. In this paper we apply Faster R-CNN to the task of company logo detection. Motivated by its weak performance on small object instances, we examine in detail both the proposal and the classification stage with respect to a wide range of object sizes. We investigate the influence of feature map resolution on the performance of those stages.Based on theoretical considerations, we introduce an improved scheme for generating anchor proposals and propose a modification to Faster R-CNN which leverages higher-resolution feature maps for small objects. We evaluate our approach on the FlickrLogos dataset improving the RPN performance from 0.52 to 0.71 (MABO) and the detection performance from 0.52 to 0.67 (mAP).

The Application of Improved YOLO V3 in Multi-Scale Target Detection

“…It is essential to determine what size of the anchor box is suitable for each scale of the network. Inspired by the method of the proposal generation [33], we used mathematical derivation based on IOU to help select the appropriate size of the anchor boxes for each scale.…”

Detection of Boulders in Side Scan Sonar Mosaics by a Neural Network

“…This deletes most of the cobbles and fine boulder information from the data if large scale mosaics were fed directly into the network. Hence, the smaller tiles exported from the backscatter mosaics were upscaled to values between 300 and 1200 pixels, which is the simplest approach to facilitate small object detection [30,31]. The size of anchor boxes used to determine the bounding box of objects were left at their standard settings of 32, 64, 128, For classification and object detection, we use an open source RetinaNet [27] implementation in Python, available on GitHub (https://github.com/fizyr/keras-retinanet, last accessed on 6 February 2019).…”

“…Therefore, it is mandatory to detect objects of the smallest possible size and to consider the minimum object size detectable by the trained models. The minimum size of objects whose detection can be trained by RetinaNet depends on a) the resolution of the input backscatter mosaic and b) the minimum anchor box of the network measured in pixels multiplied by the threshold of areal overlap of 0.5 required for a positive training [30,31]. For a minimum anchor box of 32 pixels, this results in a theoretical minimum threshold for positive training of 23 × 23 pixels.…”

“…However, the bounding box threshold criterion of 0.5 may still be missed for these size classes in case of different alignment of bounding boxes. In addition, it needs to be considered that manually digitized bounding boxes are necessarily inaccurate at the pixel-level and may not be reproduced by the model, and errors on bounding boxes effect smaller objects more than larger objects [30]. For 7 × 7 pixel bounding boxes (3.1 m 2 at 0.25 m resolution), a detection appears feasible following the upscaling, as can be observed by the increased detection frequency by the 25 m 2 model rivaling the human interpreter ( Figure 5).…”

An effective and accurate identification system of Mycobacterium tuberculosis using convolution neural networks

“…The AlexNet (Krizhevsky, Sutskever, & Hinton, 2012), VGG (Simonyan, & Zisserman, 2014), GoogLeNet (Szegedy et al, 2015), and ALL-CNN (Springenberg, Dosovitskiy, Brox, & Riedmiller, 2014) incorporated deep structures with variants of layers and have achieved remarkable performance in the classification of a large number of categories. In terms of image detection, the Faster region-based CNN (Faster R-CNN) (Ren, He, Girshick, & Sun, 2015) has been proved as an efficient detection method for small object, such as ship detection from SAR images (Kang, Leng, Lin, & Ji, 2017), company logo detection from real-world images (Eggert, Zecha, Brehm, & Lienhart, 2017), cancer cell detection (Zhang, Hu, Chen, Huang, & Guan, 2016) and gland instance detection (Xu et al, 2017) from microscopic images. Until now, relatively few studies were performed to identify M. tuberculosis from microscopic images using CNNs, except for chest X-ray TB image studies (Cao et al, 2016;Silva, Silva, Pinho, & Costa, 2017).…”