Abstract: Conventional penetration testing is destructive and therefore cannot meet the demand for rapid and nondestructive prediction of crisp persimmon firmness. Moreover, existing acoustic-vibration methods do not fully exploit the characteristics of nonstationary response signals. To address these limitations, this study developed a firmness prediction method that combines two-dimensional representations of acoustic-vibration signals with a lightweight MobileNetV3-RA model. First, 570 crisp persimmons were collected from Gongcheng County, Guangxi Zhuang Autonomous Region, China, covering four harvest batches. A sinusoidal swept-frequency excitation signal ranging from 100 to 1,500 Hz was applied to each fruit, and its acoustic-vibration response was recorded using a microphone at a sampling frequency of 8,192 Hz for 5 s. Reference firmness was measured by a texture analyzer using a penetration test with a cylindrical probe 6 mm in diameter, a penetration speed of 1 mm/s, and a penetration depth of 8 mm. Second, four methods, namely the Markov transition field (MTF), Gramian angular difference field (GADF), continuous wavelet transform (CWT), and discrete S-transform (DST), were used to convert each one-dimensional vibration response into a 256 × 256-pixel two-dimensional image. These four representations characterize the response from complementary perspectives, including temporal state transitions, angular correlations, multiscale variations, and frequency-dependent time localization, thereby allowing the suitability of each representation for firmness-related feature learning to be examined. A MobileNetV3-RA regression model was then constructed by retaining the squeeze-and-excitation (SE) mechanism of MobileNetV3 and introducing a residual attention (RA) module. The lightweight backbone was used to reduce model complexity, whereas the SE and RA modules were used to strengthen the extraction of channel and spatial features from the two-dimensional images. The samples were divided into training, validation, and prediction sets at a ratio of 6:2:2, corresponding to 342, 114, and 114 samples, respectively. A conventional two-dimensional convolutional neural network (2D-CNN) and an 18-layer residual network (ResNet18) were selected as comparison models. All models were evaluated using the same sample allocation. Model performance was assessed using the coefficient of determination for the prediction set ( R_\textp^2 ), root mean square error of prediction (RMSE_P), and residual predictive deviation (RPD_P). The results showed that CWT and DST generally provided better predictive performance than GADF and MTF, and the combination of CWT images and MobileNetV3-RA achieved the best overall result among the tested model–representation combinations. With CWT images as input, MobileNetV3-RA obtained an R_\textp^2 of 0.909, an RMSE_P of 0.885 N/mm, and an RPD_P of 3.307 on the prediction set. Under the same CWT input condition, the corresponding values for 2D-CNN were 0.843, 1.159 N/mm, and 2.526, whereas those for ResNet18 were 0.804, 1.258 N/mm, and 2.277. Compared with 2D-CNN, MobileNetV3-RA increased R_\textp^2 by 0.066 and reduced RMSE_P by 0.274 N/mm; compared with ResNet18, it increased R_\textp^2 by 0.105 and reduced RMSE_P by 0.373 N/mm. The parameter count, model storage size, floating-point operations, and mean single-sample inference time of MobileNetV3-RA were 0.80 million, 3.00 MB, 0.14 G, and 8.5 ms, respectively. These four indicators were reduced by 90.6%, 90.8%, 41.7%, and 30.3%, respectively, relative to 2D-CNN, and by 92.9%, 93.3%, 92.2%, and 45.5%, respectively, relative to ResNet18. In summary, the proposed combination of CWT-based two-dimensional acoustic-vibration representation and MobileNetV3-RA improved crisp persimmon firmness prediction while substantially reducing computational and storage requirements under the present experimental conditions. The method provides a technical reference for rapid and nondestructive fruit firmness assessment and has potential for subsequent validation in embedded devices and online sorting systems.