OBJECT RECOGNITION IN LOW LIGHT AND HAZY ENVIRONMENT

Abstract

The thesis aims to develop new methods for object recognition in low-light and hazy environment. The performance of existing object recognition methods degrades while dealing with low-light/hazy environments due to poor visibility of details and other issues in the captured images. There are two approaches to deal with low light/hazy environment: one approach to deal with low-light/hazy environment is to use enhancement/dehazing before object recognition, and the other approach focuses on direct object detection from the low-light/hazy images. The work attempted to improve object recognition using both approaches. Based on the first approach for low-light, this thesis presents two low-light image enhancement methods. The first method estimates structure-aware initial illumination from the input images based on the proposed multi-scale guided image filter. A multi-objective optimization function is formulated and solved to refine the illumination. The adaptive illumination adjust ment is developed to improve the overall lightness of the low-light images leveraging the estimated illumination. The second method develops a deep-network for simul taneous estimation of illumination and reflectance from a single image. A branched encoder-decoder architecture is developed for the decomposition task. The estimated illumination is adjusted using the deep network to improve the overall lightness. The image refinement module is developed to improve the color, contrast and other details in the enhanced images. Similarly for hazy environment, We developed a variational optimization based method for image dehazing. It is difficult to estimate depth of a scene from RGB image. We introduced a notion that the objects with same structure at a depth contains similar transmission. Thus, we developed a new method for es timation of structure-aware initial transmission leveraging the scale-adaptive bilateral filter. We formulated a new variational optimization problem with regularization terms to preserve the structural details in the final transmission. The atmospheric light is not dependent on the color; thus, we developed a uniform atmospheric light estimation method. The performance of the developed methods is compared with the various contem porary methods on a large set of images using visual and quantitative assessments. The analysis shows the superiority of the proposed methods over the existing meth ods in the enhancement/dehazing task. Furthermore, we analyzed the performance of object recognition using various enhancement methods. The enhancement method requires additional time for processing and the performance of object recognition de pends greatly on the performance of enhancement methods. Thus, we aimed to develop v a direct object recognition methods for low-light and hazy environment that do not re quire explicit enhancement to be performed before object recognition. This thesis present a new Multi-exposure refinement network for low-light object detection (MRN-LOD) to avoid the need for enhancement before recognition. The MRN-LOD contains: multi-exposure feature extractor, adaptive refinement network, and detection head. We introduced the notion of feature extraction from multi-exposure images for object recognition in low light. In addition, we proposed an adaptive re finement network to refine the features of low-light images for better recognition per formance. The detection head uses the refined features to perform object recognition. Extensive experimentation on existing real-world and synthetic datasets shows the su periority of the proposed MRN-LOD. Furthermore, The performance of the object de tection methods degrades in a hazy environment. To overcome this issue, we propose a Bi-stream feature fusion (BFF) network for object detection in a hazy environment. The BFF network consists of three modules: hybrid input, Bi-stream feature extractor (BFE), and multi-level feature fusion. We present the notion of hybrid input to extract features from the hazy images in an effective manner. The proposed BFE network extracts multi-level features from the hazy image and hybrid input. The multi-level feature fusion (MFF) network performs the convolution-based adaptive feature fusion and processes the extracted features. The proposed BFF model outperforms other state of-the-art methods in hazy environments. Another challenge in hazy object detection is the unavailability of a dataset with sufficient samples and classes. In this work, we developed a synthetic object detection dataset for a hazy environment (DHOD).