In response to the high missed detection rates and low efficiency of traditional methods in detecting tiny defects on the machining surfaces of mechanical parts, this study proposes an efficient defect detection method based on deep learning. Initially, referencing the network architectures of Resnet and Yolo, an image detection network was designed featuring a shared encoder, a classification decoder, and a localization decoder. The shared encoder is used to extract a unified feature representation; the classification decoder accomplishes efficient data classification; and the localization decoder achieves precise defect localization. Furthermore, upon acquiring high-resolution images of the machining surfaces with dimensional features, this study introduces a real-time sliding window method to perform segmented detection and classification of these images, transforming most of the target detection tasks into image classification problems, thereby further enhancing the efficiency and accuracy of defect detection and target localization. Practical results demonstrate that this method outperforms traditional approaches in terms of missed detection rates and detection efficiency, effectively addressing the challenge of detecting complex machining surface defects, and providing a high-precision, high-efficiency defect detection solution for the mechanical part machining field.