Search

Multi-object detection is one of essential prerequisites for autonomous driving under complicated traffic environments. Specially, for the cases of nighttime driving, the accuracy of object detector for those images captured by infrared cameras is inevitably affected by several unanticipated issues (such as dim and unexpected light sources), which may significantly degrade the performance of detecting multiple objects (e.g., pedestrians and vehicles). To address such problem, we propose a Light-Weighted Multi-Object Detection Scheme for Infrared Images (LW-MODS-IRI). First, we select an effective super-resolution (SR) method to enhance and repair those infrared images, which can facilitate restoring the details of images for better detection accuracy. Next, those enhanced images are trained by our LW-MODS-IRI scheme. In contrast with previous methods, we develop an adaptive computing policy for anchor clustering that aims at locating the bound of targets in a reasonable manner for higher detection precision. Moreover, a modified attention mechanism is exploited to enlarge the receptive fields for promoting the detection performance of those large objects. Furthermore, we propose an improved channel-level pruning technique to generate both compact and accurate model after the training with sparsity. Finally, the extensive experiment results on Teledyne FLIR ADAS dataset (FLIR) demonstrate that the LW-MODS-IRI scheme can have both higher inference accuracy (e.g., 10.7% more) and higher detection efficiency (e.g., 4.13 times faster) compared to the existing object detectors, which typically reveals its practical viability when it is applied to in-vehicle systems.

Image segmentation is a challenging problem in medical applications. Medical imaging has become an integral part of machine learning research, as it enables inspecting interior human body with no surgical intervention. Much research has been conducted to study brain segmentation. However, prior studies usually employ one-stage models to segment brain tissues, which could lead to a significant information loss. In this paper, we propose a multi-stage Generative Adversarial Network ( $GAN$ ) model to resolve existing issues of one-stage models. To do this, we apply a coarse-to-fine method to improve brain segmentation using a multi-stage $GAN$ . In the first stage, our model generates a coarse outline for both the background and brain tissues. Then, in the second stage, the model generates a refine outline for the white matter ( $WM$ ), gray matter ( $GM$ ), and cerebrospinal fluid ( $CSF$ ). We perform a fusion of the coarse and refine outlines to achieve high results. Despite using very limited data, we obtain an improved Dice Coefficient (DC) accuracy of up to 5% compared to one-stage models. We conclude that our model is more efficient and accurate in practice for brain segmentation of both infants and adults. In addition, we observe that our multi-stage model is 2.69–13.93 minutes faster than prior models. Moreover, our multi-stage model achieves higher performance with only a few-shot learning, in which only limited labeled data is available. Therefore, for medical images, our solution is applicable to a wide range of image segmentation applications for which convolution neural networks and one-stage methods have failed. This helps to advance the process of analyzing brain images, thus providing many advantages to the healthcare system, especially in critical health situations where urgent intervention is needed.