HiT-Det: Efficient Target Detection Based on Lightweight Network and Attention Mechanism

As the main task in computer vision, object detection is aimed at identifying the object in the image and frame its position. Recently, single-stage detection methods, which are on the basis of deep learning (such as YOLO and SSD), have been widely used in automatic driving, monitoring and other fields due to their high detection speed and real-time performance. However, the balance between accuracy and computing resources of existing mainstream detection models still needs to be optimized, especially in resource constrained scenarios. Therefore, this study proposes an efficient target detection method based on lightweight network and attention mechanism. By using lightweight backbone networks (such as MobileNetV2 and Efficient Net lite) in combination with deep separable convolution (DWConv), the amount of parameters and computation of the model is significantly reduced. At the same time, SE attention mechanism (or CBAM, coordinate attention, etc.) is introduced to dynamically adjust the weight of channel and spatial features, so as to enhance target features and restrain background interference. For the sake of optimizing feature mix, we adopt an improved feature pyramid network (FPN) strategy, which enables features of different scales to be fused more efficiently, thus improving the detection ability of small targets. Experimental results show that this method could not only prominently reduce the complexity of the model but also improve the reasoning speed while ensuring the accuracy, especially in the detection of small targets. Compared with YOLOv5s, SSD300 and other classical models, this method perform better in model volume, speed and accuracy. The experiments further demonstrate the effectiveness of the proposed lightweight design and attention mechanism.