Fast and Accurate Detection of Dim and Small Targets for Smart Micro-Light Sight

Abstract

To deal with low recognition accuracy and large time-consumption for dim, small targets in a smart micro-light sight, we propose a lightweight model DS_YOLO (dim and small target detection). We introduce the adaptive channel convolution module (ACConv) to reduce computational redundancy while maximizing the utilization of channel features. To address the misalignment problem in multi-task learning, we also design a lightweight dynamic task alignment detection head (LTD_Head), which utilizes GroupNorm to improve the performance of detection head localization and classification, and shares convolutions to make the model lightweight. Additionally, to improve the network’s capacity to detect small-scale targets while maintaining its generalization to multi-scale target detection, we extract high-resolution feature map information to establish a new detection head. Ultimately, the incorporation of the attention pyramid pooling layer (SPPFLska) enhances the model’s regression accuracy. We conduct an evaluation of the proposed algorithm DS_YOLO on four distinct datasets: CityPersons, WiderPerson, DOTA, and TinyPerson, achieving a 66.6% mAP on the CityPersons dataset, a 4.3% improvement over the original model. Meanwhile, our model reduces the parameter count by 33.3% compared to the baseline model.