Squeeze-and-excitation blocks embedded YOLO model for fast target detection under poor imaging conditions

How to detect targets under poor imaging conditions is receiving significant attention in recent years. The accuracy of object recognition position and recall rate may decrease for the classical YOLO model under poor imaging conditions because targets and their backgrounds are hard to discriminate. We proposed the improved YOLOv3 model whose basic structure of the detector is based on darknet-53, which is an accurate but efficient network for image feature extraction. Then Squeeze-and-Excitation (SE) structure is integrated after non-linearity of convolution to collect spatial and channel-wise information within local receptive fields. To accelerate inference speed, Nvidia TenorRT 6.0 is deployed into on Nvidia Jetson series low power platform. Experiments results show that the improved model may greatly achieve the inference speed without significantly reducing the detection accuracy comparing with the classic YOLOv3 model and some other up-to-date popular methods.