<p>Defect detection in plate components is essential for assessing structural integrity and minimizing industrial losses. Ultrasonic phased array inspection offers high efficiency and robust detection capabilities. However, conventional recognition systems rely on manual classification, often resulting in errors and inconsistencies. This study proposes UC-ResNet, an enhanced deep learning framework for classifying ultrasonic phased array images of plate components. The improved architecture employs a U-Net structure with three-stage encoding-decoding and skip connections to extract and fuse multi-scale defect features from ultrasonic images while suppressing artifacts. The model incorporates a Convolutional Block Attention Module that reconstructs feature maps through dual-path channel-spatial weight allocation, enhancing responses to critical defect features, while its coordinate attention mechanism improves bounding box localization accuracy. OpenCV-based image enhancement, including grayscale conversion and morphological filtering, prepares ultrasonic data for feature extraction. Experimental results demonstrate that UC-ResNet on a dedicated ultrasonic dataset achieves a mean average precision (mAP50) of 92.9% and an accuracy of 96.3%, outperforming the original ResNet by 2.2%. These findings validate the effectiveness of the proposed approach for automatic defect identification in ultrasonic phased array imaging.</p>