1, a small share of the target pixel problem
One is the relative size, length and width dimensions as the target image size of the original 0.1, can be considered a small target;
the other is defined absolute dimensions, i.e., smaller than the target size of 32 * 32 pixels can be considered to be small aims.
Summary of small target is an image area covering a smaller class of target compared with the conventional target, small target small amount of information, the training data is difficult to mark, which results in good general target detection method for detecting small effects on the target, and small target detection method designed specifically for the high complexity or often do not have the versatility. based on the present target detection method based on a multi-scale flash zone convolutional neural network oriented small target (faster-regions with convolutional neural network, Faster-RCNN ) detection algorithm. the characteristics of a convolutional neural network, the network structure is modified Faster-RCNN, so that the network may use high-level and low-level features of the multi-scale object detection Meanwhile, in order to enhance the low-level feature for the accuracy of small target detection task main test basis. Meanwhile, for the training data is difficult to mark, using data obtained from the search engine used to train the model. because these different training data with the task of test data distribution, and use of down-sampling and the method of sampling of the target high-resolution training image transformation, so the training images The feature of the test image more similar to the distribution of the experimental results show that: the proposed method on a small target detection precision tasks mean average (mean average precision, mAP) can be increased from about 5% higher than the original Faster-RCNN.
http://crad.ict.ac.cn/CN/abstract/abstract3866.shtml
Faster-RCNN multiscale algorithm for small target detection
faster-rcnn principle Introduction - CSDN blog - Lin_xiaoyi's blog https://blog.csdn.net/Lin_xiaoyi/article/details/78214874
2, the recognition rate algorithm
Deep Learning Lecture 24: Overview of computer vision, object detection algorithm - Cloud + Community - Tencent cloud https://cloud.tencent.com/developer/news/302362
