MOD-YOLO: Rethinking the YOLO architecture at the level of feature information and applying it to crack detection

Civil infrastructure plays an important role in daily life. If cracks are not found in time, they may cause immeasurable losses to people and property. Therefore, timely and accurate detection and localization of cracks is of great significance. Considering the possible loss of channel information and the lack of receptive field in the previous You Only Look Once (YOLO) series of algorithms, we design a Maintaining the Original Dimension-YOLO (MOD-YOLO) algorithm and apply it to crack detection in civil infrastructure. All the improved schemes in the algorithm are plug-and-play. First, we propose Maintaining the Original information-Deeply Separable Convolution (MODSConv), which solves the problem that we cannot interact with information between channels in the original feature layer (as seen in classical deeply separable convolutions). Second, we propose Global Receptive Field-Space Pooling Pyramid-Fast to obtain global perspective information and mitigate the impact of different scales. Third, Distinctive and Average Features-Coordinate Attention (DAF-CA) is proposed. This not only deals with the reference average information but also considers salient information. With this, we can find and enhance key information more accurately. In addition, we design Maintaining the Original information-Deeply Separable Layer (MODSLayer), which protects the rich information between channels in a way that does not reduce the dimensionality of the channel. At the same time, MODSLayer builds the backbone and neck of the network. The network is named Maintaining the Original information-Deeply Separable Network(MODSNet). Finally, Maintaining the Original Dimension Light-Head is designed for channel non-dimensionality reduction. It maintains as much feature layer information as possible before prediction under the premise of being as lightweight as possible, which significantly improves detection accuracy and detection speed. The experimental results show that our algorithm improves the accuracy by 27.5% to 91.1% on the crack dataset compared to the YOLOX algorithm with the crack detection time basically the same as the YOLOX algorithm, and with the parameter amount reduced by 19.7% and the computational complexity reduced by 35.9%. Meanwhile, experiments on COCO2017, VOC2007 and other datasets verify its good generalizability. The whole vehicle deployment scheme for crack detection is proposed and used to implement the algorithm to detect cracks while the vehicle is moving, and the accompanying experiments prove that our algorithm is able to complete the task of detecting cracks while the vehicle is moving very well.