Traditionally, the sorting of small mechanical parts has heavily relied on artificial vision, which is prone to false or missed detection. The emergence of deep learning-based object detection algorithms offers promise to solve this problem. However, the small size of mechanical parts makes it challenging to achieve high accuracy. To address this issue, an improved object detection algorithm based on YOLOv7 is proposed. Firstly, the CBAM attention module is incorporated into the YOLOv7 to emphasize object features and suppress secondary information. Furthermore, the YOLOv7 is integrated with the switchable atrous convolution (SAC) to enhance the feature extraction capability. SAC allows the model to select the outputs of the convolution adaptively. Finally, to realize a more precise localization of the proposed algorithm, the CIoU loss function is replaced with the Wise-IoU v3 (WIoU v3) loss function. The proposed algorithm’s effectiveness is substantiated through extensive experimental evaluations and achieves a remarkable mean Average Precision (mAP) score of 96%, surpassing the original YOLOv7 by 18.6%. The experiment results demonstrate the excellent performance of the proposed algorithm.