Design of visual detection and location system for software code defects based on deep learning and computer vision

Under the background of continuous iteration of large-scale software system, the complexity of source code structure is increasing, and the traditional defect detection methods based on rules or semantic parsing have obvious shortcomings in cross-language adaptation and defect location accuracy. Focusing on this problem, this paper proposes a visual detection and location method of software code defects based on deep learning and computer vision. In this paper, the source code character sequence is mapped into a fixed-size gray-scale image, and the information such as code typesetting, indentation level and structural density is preserved, and a collaborative detection framework of convolutional neural network and attention model is constructed to realize the joint modeling of defect identification and line-level positioning. Experiments are carried out on multilingual open source data sets. The comparative results show that this method is superior to the single model structure in accuracy, recall and positioning error control, and maintains stable performance under cross-data set conditions. The visual heat map results provide an intuitive basis for defect determination and help to improve the understandability of the analysis process. The research results show that visual modeling of code provides an effective technical path for software defect detection, and has reference value for automatic code analysis in complex engineering environment.