Abstract: With the rapid development of the internet and digital technology, various types of data are being generated explosively. Image recognition tasks urgently require efficient analysis and processing techniques, thus driving the rapid development of artificial intelligence technology. The existing conventional image recognition technology based on the von Neumann architecture is significantly limited in computation speed and power consumption due to the separation of sensing, storage, and computation units. To address this issue, a novel on-chip image recognition solution with a non-von Neumann architecture has been developed. This new image recognition technology, based on in-memory computing with artificial synapses, offers lower power consumption and faster processing speeds compared to traditional von Neumann architecture-based image recognition technology. However, due to limitations in the design and manufacturing technology of synaptic components, the conductance of artificial synapses is often discrete, leading to scattered weight values in such new image recognition technologies. This means that it is impossible to deploy ideal neural network weights, thereby reducing the effectiveness of image recognition. To solve this problem, this study designed related algorithms using the Keras library, quantizing continuous ideal weights into discrete weights to simulate the non-ideal characteristics of actual devices. The study explored the effects of weight dispersion, distribution patterns of weights, and different image regions on the accuracy of neural networks in image recognition tasks. The research found that increasing the number of weight options, i.e., using higher precision devices, can effectively reduce the impact of non-ideal weight characteristics; the discrete nature of weights has a greater impact on image regions with high information density, and these areas should be prioritized when resources constrains; different distribution patterns of weights produce varying effects in different image regions, and each sub-region has its optimal distribution. Using an "integrated distribution model" composed of the optimal weight distributions for different sub-regions can most effectively reduce the impacts brought by discrete characteristics, although the performance improvement is limited compared to a single distribution pattern. The results reveal the specific impacts of different weight quantization levels in various image areas on recognition performance, providing valuable references for optimizing image recognition systems based on artificial synapses.