A lightweight object detection network in low-light conditions based on depthwise separable pyramid network and attention mechanism on embedded platforms

Abstract

Deep learning-based object detection algorithms have been widely used and are effective in autonomous driving. But their performance degrades dramatically under low-light conditions. The existing mitigation approach for low-light object detection uses image enhancement preprocessing to improve detection performance with limited success. In addition, this preprocessing approach incurs extra computation costs, consumes more resources, and makes it more challenging to implement the algorithm on embedded platforms. This paper proposes a novel object detection network in low-light conditions for embedded platforms. The new system consists of two innovative modules, including a lightweight low-light enhancement network DS-PyLENet and an anchor-free lightweight object detector CFEDet. DS-PyLENet is a three-level pyramid network configured with depthwise separable convolution residual blocks (DSCRBs) and multiscale DSCRBs. The CFEDet is configured with an improved EfficientNet backbone network with Coordinate Attention (CA) and vanilla MBconv, i.e., CA-fused EfficientNet, an improved Path Aggregation Network (PAN) fusion module, and an anchor-free Global Focal Loss (GFL) detection head. Two evaluation metrics, relative recall ratios, are proposed to depict the effectiveness of image enhancement on object detection more intuitively. The new model is demonstrated to be superior to the assembled model of EnlightenGAN and YOLOv5n in terms of detection accuracy and speed on GPU and embedded platforms. The testing results of the new model reveal that 83.5% mAP is achieved on the Exdark dataset and 67.4% mAP on the DarkFace dataset. The running rates of GTX 1080Ti and NVIDIA Jetson Xavier NX are 22 FPS and 2.6 FPS, respectively. The recall of CFEDet with DS-PyLENet increased to 22.1% and significantly improved over 18% with Zero-DCE and 17.2% with EnlightenGAN.