A comparative study of different deep learning approaches to object detection

Abstract

Deep learning typically necessitates substantial datasets for effective model training, yet the collection and annotation of such data are often time - consuming and challenging. In the realm of road surface defect detection, traditional methods predominantly rely on single crack - like texture features for defect identification. Nevertheless, these approaches frequently encounter limitations when dealing with complex road conditions, as they require extensive crack datasets that are difficult to acquire, particularly under extreme circumstances. Consequently, they may fail to effectively recognize other types of road surface defects. Besides cracks, common road defects such as potholes are widespread on highway surfaces, demanding a more comprehensive detection method. Moreover, real - world road scenarios are often complicated by factors like debris, tree shadows, and passing vehicles, which can interfere with detection. Traditional defect detection methods are prone to misinterpretations and omissions, leading to low identification efficiency and suboptimal outcomes. Relying solely on manual assessment for road condition evaluation is not adequate to meet the requirements for high - quality road maintenance, thus highlighting the urgent need for more efficient detection methods. To slove the issues to be more accurate, demanding datasets, and high computational costs associated with traditional road surface defect detection algorithms, improvements were made based on the existing YOLOv8-ip object detection algorithm, proposing a new algorithm called YOLOv8-ip . To enhance accuracy, a coordination attention mechanism was introduced. This mechanism perceives the input tensor in both horizontal and vertical directions simultaneously, emphasizing the importance of different channel information and encoding spatial information. It enables the model to more accurately locate and identify objects of interest, thereby improving object detection accuracy while maintaining computational speed. To increase object detection precision without compromising computational speed, two main operations were also performed. First, the original C2f modules was replaced with C2fGhost modules, optimizing the floating - point operations in the feature channel fusion process. This resulted in reduced model parameters, enhanced feature representation performance, and decreased computational load. Second, the original SPPF module was introduced in the model with the SimSPPF module, which improved computational speed by replacing the activation function. Ultimately, the experimental detection was conducted on road defect images using the YOLOv8-ip model and compared it with existing excellent object detection algorithms. Compared to previous algorithms, the method achieved a 1.9% improvement in mAP@0.5 and an 8.65% reduction in computational load, realizing a dual enhancement in accuracy and efficiency. Keywords: pavement defects, deep learning, yolo, neural network