Image Pre-processing Techniques for Improving Object Recognition Performance of Deep Neural Networks

Abstract

In this research, image pre-processing techniques like tone mapping and dehazing have been proposed for intelligent traffic sign recognition. An innovative Intelligent Traffic Sign Recognition (ITSR) system with the illumination pre-processing capability has been developed successfully to detect signs in dark region of an image. New Dark Area Sensitive Tone Mapping (DASTM) technique can enhance the illumination of only dark regions of an image with minor impact on bright regions, which has been used as a pre-processing module for this novel traffic sign recognition system. DASTM has been combined with a TS detector for the detection of three classes of traffic signs by using an optimized version of YOLOv3. ITSR has been trained for prohibitory, mandatory, and danger classes on a dataset of Korean traffic signs. Thus, 90.07% Mean Average Precision (MAP) has been achieved for challenging Korean Traffic Sign Detection (KTSD) dataset which has never been achieved earlier (previous best result was 86.61% for KTSD). Furthermore, 100% on German Traffic Sign Detection Benchmark (GTSDB). Results comparisons of ITSR with latest D-Patches, TS detector, and YOLOv3 show that new ITSR significantly outperforms other methods in recognition performance. Features of traffic signs on road images become dull due to low visibility in hazy weather. This research also describes effect of image dehazing on deep learning based traffic sign recognition. Reliability guided fusion scheme for image dehazing has been merged with Traffic Sign (TS) detector to recognize the traffic signs from hazy day road images. Dehazing has been applied on hazy images and then detection algorithm is used for the identification of three traffic signs classes. Experimental results show 6.77% increase in detection accuracy of TS detector on hazy images dataset, owing to the dehazing.