IMAGE PREPROCESSING ALGORITHM TO REDUCE THE PROBABILITY OF OVERFITTING OF CONVOLUTIONAL NEURAL NETWORKS ON A NEURAL ACCELERATOR
Abstract
The main volume of requirements in early detection systems are imposed on the performance of digital image processing algorithms that are implemented on embedded devices with limited computing resources. In the early detection problem, objects in images are represented by a small number of pixels. Therefore, in order to ensure the required accuracy characteristics of the algorithms for searching and recognizing objects in images, algorithms for preliminary processing of a sequence of video frames are used to expand the original feature space. Processing high-resolution images by image preliminary processing algorithms leads to an unacceptable time delay in the execution of the algorithm and is a "bottleneck" of the entire algorithm. In this paper, an algorithm for preliminary processing of a sequence of video frames for a neural accelerator is proposed in order to expand the feature space, which allows increasing the speed of data processing. This is achieved by merging the image preliminary processing algorithm with the feature extractor of a convolutional neural network and transferring the execution of a new feature extractor to the computing power of the neural accelerator. The developed algorithm was tested by conducting a computational experiment. On NVIDIA Jetson and Rockchip computing devices, the algorithm of preliminary processing is implemented twice on the central processor and neural accelerator, according to the developed algorithm. Estimates of the execution time of the algorithms are obtained, which show that the proposed algorithm of preliminary image processing for the neural accelerator allows increasing the data processing speed by 1.4 - 4 times depending on the type of bit depth of calculations. However, the transition to the integer type of calculations of the CNN model with a modified feature extractor leads to a decrease in the Mean Average Precision metric by 5–19.4%, characterizing the integral average accuracy of searching and recognizing objects in images
