In this work, six convolutional neural networks (CNNs) have been trained based on 15638 superflare candidates on solar-type stars, which are collected from the three years of Transiting Exoplanet Survey Satellite (TESS) observations. These networks are used to replace the manually visual inspection, which was a direct way of searching for superflares, and exclude false-positive events in recent years. Unlike other methods, which only used stellar light curves to search for superflare signals, we try to identify superflares through TESS pixel-level data with lower risk of mixing false-positive events and give more reliable identification results for statistical analysis. The evaluated accuracy of each network is around 95.57%. After applying ensemble learning to these networks, the stacking method promotes accuracy to 97.62% with a 100% classification rate, and the voting method promotes accuracy to 99.42% with a relatively lower classification rate at 92.19%. We find that superflare candidates with short duration and low peak amplitude have lower identification precision, as their superflare features are hard to be identified. The database includes 71732 solar-type stars and 15,638 superflare candidates from TESS with corresponding feature images and arrays, and the trained CNNs in this work are public available.