果蝇与脊椎动物的视觉系统在结构与作用模式上高度相似，因此果蝇常被作为研究视觉系统的模式生物。果蝇视觉系统包括以下几个部分：光感受器，薄板，髓质，小叶区，小叶板。现有研究发现尽管果蝇小眼间夹角的最小角为4.5度，但果蝇通过微眼动使其视觉实际最小可分辨角低于4.5度，甚至可达到1.16度，因此认为果蝇产生了高于复眼分辨率的视觉能力，即视觉超敏性。视觉超敏性理论认为果蝇光感受器采取微扫视采样策略最大化了对外界信息的采样。同时，跨越薄板与髓质的L2神经元接收来自光感受器的超敏信息，并对其进行编码处理后，将信息传递给下游神经元。

果蝇此类采样编码策略能够帮助果蝇提高对物体的识别能力。同时果蝇在光学系统分辨率有限情况下，其神经元结构能够提取来自外界的视觉信息并且编码特征集。受到该采样与编码策略启发，并结合目前对视觉数据进行识别、理解、并且运用的需求。本研究设计了基于果蝇早期视觉神经通路神经网络模型，用于模拟果蝇光感受器对信息进行采样以及L2神经元编码过程。进一步验证果蝇光感受器采样形式及果蝇薄板中L2神经元连接关系。

本研究在长短记忆神经网络模型的基础上，设计了果蝇早期视觉系统的简化模型。其中模型的输入为以时间序列构成的运动光栅视觉刺激，即果蝇所观察的视觉刺激图形，输出为L2神经末梢活动。模型训练使用的L2神经元末梢活动数据集，来自18只果蝇在不同视觉刺激下，使用双光子钙离子成像技术测量的L2神经元末梢活动数据。

本研究模型分为两层，第一层为光感受器模型，参考小眼中光感受器数量和光感受器采样方式，采用不同的过滤器模拟果蝇光感受器细胞对视觉刺激采样方式。第二层为薄板中L2神经元连接组模型，参考L2神经元在薄板中与同一小眼中L4神经元以及相邻小眼L4神经元、L2神经元的连接关系，以及L2神经元末梢数量，选取不同权重矩阵为循环内核对L2神经元在薄板中的连接组进行模拟。

本研究实验结果证明，果蝇在光感受器通过微扫视策略使复眼产生高斯采样，在薄板中L2神经元连接组为L2神经元与同一小眼L4神经元以及相邻小眼中L4神经元互相联系，接收来自这些神经元的反馈信息。并且模型观察到与过往研究一致的现象，果蝇L2神经末梢活动对视觉刺激的方向具有偏好性。

此模型可运用于计算机视觉中，以优化采集图像分辨率。

The visual system of Drosophila is highly similar in structure and mode of action to that of vertebrates, so Drosophila is often used as a model organism for studying the visual system. The Drosophila visual system consists of the following components: photoreceptors, lamina, medulla, lobular, and lobular plate. Existing studies have found that although the smallest angle of the interommatidial angle in Drosophila is 4.5 degrees, the actual smallest resolvable angle in Drosophila vision is lower than 4.5 degrees and can even be as high as 1.16 degrees during saccadic behaviors, and Drosophila is thought to produce acuity above the resolution of the compound eye. The theory of hyperacute vision suggests that Drosophila photoreceptors adopt a microsaccadic sampling strategy to maximize photoreceptors’encoding capacity. At the same time, L2 neurons spanning the lamina and medulla receive the hyperacute information from the photoreceptors, encode it, and then transmit it to downstream neurons.

Such sampling and encoding strategies in Drosophila can help Drosophila to improve its acuity. At the same time, the neuronal structure of Drosophila can extract visual information from the outside world and encode feature sets under the limited resolution of the optical system. Inspired by this sampling and encoding strategy, and combined with the current demand for recognizing, understanding, and applying visual data, this study was designed based on Drosophila's early sampling and encoding strategy. In this study, we designed a neural network model based on Drosophila early visual neural pathways to simulate the sampling of information by Drosophila photoreceptors and the encoding process of L2 neurons. We further validated the form of Drosophila photoreceptor sampling and the L2 neuron connectome in the lamina.

In this study, a simplified model of the Drosophila early visual pathway was designed based on the long and short memory neural network model. The input of the model is made of dynamically narrowing bar gratings，which is the visual stimulus graph observed by Drosophila, and the output is the L2s terminal activity. The L2 neuron terminal activity dataset used for model training was derived from L2 neuron terminal activity data measured using two-photon calcium imaging in 18 fruit flies under different visual stimulus.

The model consists two layers. The first layer is the photoreceptor model, combining the number of photoreceptors and photoreceptor connections in ommatidium, using different filters to simulate the way Drosophila photoreceptor cells sample visual stimulus. The second layer is the L2 neuron connectome model in the lamina, regarding the connections between L2 neurons in the lamina and L4 neurons in the same ommatidium and L4 neurons and L2 neurons in adjacent ommatidia, and different weight matrices are selected as the cyclic kernel to simulate the connections of L2 neurons in the lamina.

The experimental results of this study demonstrate that Drosophila produces Gaussian sampling in the compound eye by a microsaccadic behaviors in photoreceptors, and the L2 neuron connectome in lamina are L2 neurons interconnected with L4 neurons and L4 neurons in adjacent ommatidia. And the model observes that consistent with previous studies, Drosophila L2s terminal activity has a preference for the direction of visual stimulus.

This model can be applied to computer vision to optimize the resolution of acquired images.