The texture feature extraction including grayscale co-occurrence matrix and various shape feature extraction methods are adopted in this paper, as well as convolutional neural network based on Visual Geometry Group-16 structure. In particular, the Squeeze-and-Excitation module and dilated convolution technique are introduced to improve the model, aiming to enhance its feature extraction and classification capabilities. On the JPEGWELD dataset, the improved model had 98.7% accuracy in the training set, 97.9% accuracy in the test set, and 98.7% recall rate. In the comparative analysis, although the number of parameters of the improved VGG16 model was 33.64M and the maximum model size was 385MB, the detection time was only 1.3s. The results demonstrated that the model had efficient optimization and computational performance, with a good balance between design and optimization while maintaining a short detection time. The proposed method exhibits high accuracy and efficiency in the detection of various types of weld defects, demonstrating strong universality and adaptability. Its applicability to diverse industrial settings is evident. The study provides an effective solution for industrial automated inspection, which is of great significance to improve the quality control level and production efficiency of manufacturing industry.