Due to factors involved in the manufacturing process, aluminum alloy materials are prone to various internal welding defects, such as pores, slag inclusion, and incomplete penetration. However, in the DR (digital radiography) image defect detection of aluminum alloy welds, detection accuracy of the model remains insufficient. Thus far, defect detection in DR images is generally determined and located manually. However, manual film evaluation involves a high workload, with low efficiency and other issues such as false and missed detection. With the rapid development of digital image processing technology, deep learning has been widely applied for object recognition. This study proposes a lightweight YOLOv7Tiny based weld defect detection model, YOLOv7TS, to realize DR image defects detection of aluminum alloy welds.

First, a TSCODE decoupling head was added to improve the algorithm’s ability to detect small targets. To address the high aspect ratio of incomplete penetration defects and low recall rate, the Upsampling operator was changed to CARAFE to improve the receptive field. Second, for small pixel defects such as pores and slag inclusion, an SPD-Conv convolutional layer was added to enhance the small target detection ability of the model. Finally, a SimAM attention mechanism was added to reduce the depth and width of the model and to improve the overall model performance and ELAN layer.

For pore, slag inclusion, and incomplete penetration, the average precision (AP) of the YOLOv7TS model reached 89.9%, 94.2%, and 96.3%, respectively. Compared with the original YOLOv7Tiny model, average accuracy increased by 8.2, 3.7, and 2.2 percentage points, and the overall accuracy was compared to the original model, mAP@0.5, improved by 4.6 percentage points (Table 1). Meanwhile, the model parameter quantity decreased by 5% compared to the original model. Although the FPS index decreased from 222 to 208, it still meets the target detection speed requirements (Table 2).

This study focuses on key challenges including low accuracy and large model parameters for incomplete penetration defect detection in aluminum alloy weld DR images using the YOLO model. To address these challenges, we improved the YOLOv7Tiny model and proposed a new model: YOLOv7TS. The proposed model effectively improves weld defect detection accuracy. First, the addition of a TSCODE decoupling head increases the average accuracy, however, this increases the number of parameters. Second, by replacing the Upsampling operator with CARAFE and increasing the model receptive field, the average accuracy is improved. Subsequently, the first-layer convolution module is replaced with the SPD-Conv module, and a SimAM attention mechanism is included in the ELAN module. The depth and width of the model were reduced to one-third and half that of the original model, resulting in an average accuracy improvement of 4.6 percentage points and 5% decrease in parameter quantity compared to the original model. Furthermore, the proposed YOLOv7TS model demonstrates higher detection accuracy and smaller parameter size, making it more straightforward to deploy to other terminal devices.