Performance Improvement of Single-stage Target Detection by Tuning Parameters in CNN

Abstract

Single-stage Target detection can find almost all the objects in the images and determine the category and location of the objects, which includes two key points: object positioning and object classifying. Deep learning methods have been widely used to target detection. However, parameters are tuned to find relations between parameters and performance of CNN, including training time needed, detecting time, and detection precision. This paper studies the influence of the parameters in CNN on the accuracy of single-stage target recognition and improves the performance of single-stage target detection. First, the data comes from PASCAL VOC 2012. Then, a series of related parameters are analyzed in our CNN model, including image size and batch size, the Strides, the Anchor size and number, etc. Finally, by changing the single parameter and keeping other parameters the same, we study how these parameters influence the performance of our model. The results show that the image size affects the time and accuracy of the training process, and there is an antagonistic relationship between them. Besides, in the training process, we find that with the training time increasing the training loss decreased. And if reasonably increasing the Strides, the Input shape and the Non-Maximum Suppression (NMS) threshold can improve the recognition rate. However, if decreasing the Anchor size and number and the Strides greatly can increase the running time. Besides the Anchor size and number is having a certain connection related to the Mean Average Precision (mAP), and the NMS has a certain positive impact on the MAP. Our analysis demonstrates that setting reasonable NMS, strides, image size and the number of anchors which belong to Convolutional Neural Network (CNN) can make the precision rate higher. Simultaneously, it can accelerate the training process.