Large Receptive Fields Research Articles

FCPFNet: Feature Complementation Network with Pyramid Fusion for Semantic Segmentation

Traditional pyramid pooling modules have shown effective improvements in semantic segmentation tasks by capturing multi-scale feature information. However, their limitations arise from the shallow structure, which fails to fully extract contextual information, and the fused multi-scale feature information lacks distinctiveness, resulting in issues with the final segmentation discriminability. To address these issues, we proposes an effective solution called FCPFNet, which is based on global contextual prior for deep feature extraction of detailed information. Specifically, we introduce a novel deep feature aggregation module to extract semantic information from the output feature map of each layer through a deep aggregation of context information module, and expands the effective perception range. Additionally, we propose an Efficient Pyramid Pooling Module (EPPM) to capture distinctive features through communicating information between different sub-features and performs multi-scale fusion, which is integrated as a branch within the network to complement the information loss resulting from downsampling operations. Furthermore, in order to ensure the richness of image detail feature information and maintain a large receptive field to obtain more contextual information, EPPM concatenates the input feature map and the output feature map of the pyramid pooling module to acquire more comprehensive global contextual information. It has been demonstrated by experiment that the method described in this article achieves competitive performance on the challenging scene segmentation datasets Pascal VOC 2012, Cityscapes and Coco-Stuff, with MIOU of 81.0%, 78.8% and 40.1%, respectively.

CDYL for infrared and visible light image dense small object detection

To address the phenomenon of many small and hard-to-detect objects in infrared and visible light images, we propose an object detection algorithm CDYL (Convolution to Fully Connect-ed-Deformable Convolution You only Look once) based on the CFC-DC (Convolution to Fully Connected-Deformable Convolution) module. The core operator of CDYL is CFC-DC, making our model not only have a large effective receptive field in infrared and visible light images, but also have adaptive spatial aggregation conditioned by input and task information. As a result, the CDYL reduces the strict inductive bias of traditional CNNs and has long-range dependence for large kernel convolution as well as adaptive spatial aggregation, deeply mining of edge and correlation information in images to enhance sensitivity to small objects, thereby improving performance in dense small object detection tasks. In order to improve the ability of the CFC-DC module to perceive the detailed information of the image, we use the Mish activation function, which has a wider minima which improves the generalization. The effectiveness as well as the generalization of CDYL is evaluated on an infrared image dataset and an UAV image dataset, and it is compared with other state-of-the-art object detection algorithms. Compared to the baseline network YOLOv8l, our model achieved a 3.0% improvement in mAP0.5 in infrared image detection tasks and a 1.1% improvement in mAP0.5 in visible light image detection tasks. The experimental results show that the proposed algorithm achieves superior average precision values on both infrared and visible light images, while maintaining a light weight. Code is publicly available at https://github.com/yangzhu1/CDYL.

A multi convolution pooling group fault diagnosis model with high generalization across data sets and large receptive field characteristics considering industrial environmental noise

A multi convolution pooling group fault diagnosis model with high generalization across data sets and large receptive field characteristics considering industrial environmental noise

Automatic Identification and Severity Classification of Retinal Biomarkers in SD-OCT Using Dilated Depthwise Separable Convolution ResNet with SVM Classifier

Purpose Diagnosis of Uveitic Macular Edema (UME) using Spectral Domain OCT (SD-OCT) is a promising method for early detection and monitoring of sight-threatening visual impairment. Viewing multiple B-scans and identifying biomarkers is challenging and time-consuming for clinical practitioners. To overcome these challenges, this paper proposes an image classification hybrid framework for predicting the presence of biomarkers such as intraretinal cysts (IRC), hyperreflective foci (HRF), hard exudates (HE) and neurosensory detachment (NSD) in OCT B-scans along with their severity. Methods A dataset of 10880 B-scans from 85 Uveitic patients is collected and graded by two board-certified ophthalmologists for the presence of biomarkers. A novel image classification framework, Dilated Depthwise Separable Convolution ResNet (DDSC-RN) with SVM classifier, is developed to achieve network compression with a larger receptive field that captures both low and high-level features of the biomarkers without loss of classification accuracy. The severity level of each biomarker is predicted from the feature map, extracted by the proposed DDSC-RN network. Results The proposed hybrid model is evaluated using ground truth labels from the hospital. The deep learning model initially, identified the presence of biomarkers in B-scans. It achieved an overall accuracy of 98.64%, which is comparable to the performance of other state-of-the-art models, such as DRN-C-42 and ResNet-34. The SVM classifier then predicted the severity of each biomarker, achieving an overall accuracy of 89.3%. Conclusions A new hybrid model accurately identifies four retinal biomarkers on a tissue map and predicts their severity. The model outperforms other methods for identifying multiple biomarkers in complex OCT B-scans. This helps clinicians to screen multiple B-scans of UME more effectively, leading to better treatment outcomes.

MSFANet: multi-scale fusion attention network for mangrove remote sensing lmage segmentation using pattern recognition

Mangroves are ecosystems that grow in the intertidal areas of coastal zones, playing crucial ecological roles and possessing unique economic and social values. They have garnered significant attention and research interest. Semantic segmentation of mangroves is a fundamental step for further investigations. However, mangrove remote sensing images often have large dimensions, with a substantial portion of the image containing mangrove features. Deep learning convolutional kernels may lead to inadequate receptive fields for accurate mangrove recognition. In mangrove remote sensing images, various challenges arise, including the presence of small and intricate details aside from the mangrove regions, which intensify the segmentation complexity. To address these issues, this paper primarily focuses on two key aspects: first, the exploration of methods to achieve a large receptive field, and second, the fusion of multi-scale information. To this end, we propose the Multi-Scale Fusion Attention Network (MSFANet), which incorporates a multi-scale network structure with a large receptive field for feature fusion. We emphasize preserving spatial information by integrating spatial data across different scales, employing separable convolutions to reduce computational complexity. Additionally, we introduce an Attention Fusion Module (AFM). This module helps mitigate the influence of irrelevant information and enhances segmentation quality. To retain more semantic information, this paper introduces a dual channel approach for information extraction through the deep structure of ResNet. We fuse features using the Feature Fusion Module (FFM) to combine both semantic and spatial information for the final output, further enhancing segmentation accuracy. In this study, a total of 230 images with dimensions of 768 pixels in width and height were selected for this experiment, with 184 images used for training and 46 images for validation. Experimental results demonstrate that our proposed method achieves excellent segmentation results on a small sample dataset of remote-sensing images, with significant practical value. This paper primarily focuses on three key aspects: the generation of mangrove datasets, the preprocessing of mangrove data, and the design and training of models. The primary contribution of this paper lies in the development of an effective approach for multi-scale information fusion and advanced feature preservation, providing a novel solution for mangrove remote sensing image segmentation tasks. The best Mean Intersection over Union (MIoU) achieved on the mangrove dataset is 86%, surpassing other existing models by a significant margin.

Weakly supervised medical image registration with multi-information guidance

In recent years, the registration method based on deep learning has received extensive attention from scholars due to its superiority in real-time performance. Most of the work directly use convolutional neural networks (CNNs) to map the image to be registered into the transform space. However, the receptive field of CNNs is limited, and multi-layer convolution superposition is needed to obtain a relatively large receptive field. Transformer-based methods can better express spatial relationships through attention mechanisms. However, the self-attention and the multi-head mechanisms make each small block calculate the relationship with other small blocks regardless of distance. Due to the limited moving range of corresponding voxel points in the medical images, this long-distance dependence may cause the model to be interfered by long-distance voxels. In this paper, we convert the spatial transformation of the corresponding voxel points into the calculation of the basic vector basis to propose the SV-basis module and design a two-stage multi-scale registration model. Experiments are carried out on brain and lung datasets to prove the effectiveness and universality of the proposed registration method. According to the anatomical characteristics of medical images, the corresponding loss function is designed to introduce mask information into the registration task. The experimental results show that the proposed method can accurately register brain and lung images.

ICNet: A Dual-Branch Instance Segmentation Network for High-Precision Pig Counting

A clear understanding of the number of pigs plays a crucial role in breeding management. Computer vision technology possesses several advantages, as it is harmless and labour-saving compared to traditional counting methods. Nevertheless, the existing methods still face some challenges, such as: (1) the lack of a substantial high-precision pig-counting dataset; (2) creating a dataset for instance segmentation can be time-consuming and labor-intensive; (3) interactive occlusion and overlapping always lead to incorrect recognition of pigs; (4) existing methods for counting such as object detection have limited accuracy. To address the issues of dataset scarcity and labor-intensive manual labeling, we make a semi-auto instance labeling tool (SAI) to help us to produce a high-precision pig counting dataset named Count1200 including 1220 images and 25,762 instances. The speed at which we make labels far exceeds the speed of manual annotation. A concise and efficient instance segmentation model built upon several novel modules, referred to as the Instances Counting Network (ICNet), is proposed in this paper for pig counting. ICNet is a dual-branch model ingeniously formed of a combination of several layers, which is named the Parallel Deformable Convolutions Layer (PDCL), which is trained from scratch and primarily composed of a couple of parallel deformable convolution blocks (PDCBs). We effectively leverage the characteristic of modeling long-range sequences to build our basic block and compute layer. Along with the benefits of a large effective receptive field, PDCL achieves a better performance for multi-scale objects. In the trade-off between computational resources and performance, ICNet demonstrates excellent performance and surpasses other models in Count1200, AP of 71.4% and AP50 of 95.7% are obtained in our experiments. This work provides inspiration for the rapid creation of high-precision datasets and proposes an accurate approach to pig counting.

AML‐Net: Attention‐based multi‐scale lightweight model for brain tumour segmentation in internet of medical things

AbstractBrain tumour segmentation employing MRI images is important for disease diagnosis, monitoring, and treatment planning. Till now, many encoder‐decoder architectures have been developed for this purpose, with U‐Net being the most extensively utilised. However, these architectures require a lot of parameters to train and have a semantic gap. Some work tried to make a lightweight model and do channel pruning that made a small receptive field which compromised the accuracy. The authors propose an attention‐based multi‐scale lightweight model called AML‐Net in Internet of Medical Things to overcome the above issues. This model consists of three small encoder‐decoder architectures that are trained with different scale input images along with previously learned features to diminish the loss. Moreover, the authors designed an attention module which replaced the traditional skip connection. For the attention module, six different experiments were conducted, from which dilated convolution with spatial attention performed well. This attention module has three dilated convolutions which make a relatively large receptive field followed by spatial attention to extract global context from encoder low‐level features. Then these fine features are combined with the decoder's same layer of high‐level features. The authors perform the experiment on a low‐grade‐glioma dataset provided by the Cancer Genome Atlas which has at least Fluid‐Attenuated Inversion Recovery modality. The proposed model has 1/43.4, 1/30.3, 1/28.5, 1/20.2 and 1/16.7 fewer parameters than Z‐Net, U‐Net, Double U‐Net, BCDU‐Net and CU‐Net respectively. Moreover, the authors’ model gives results with IoU = 0.834, F1‐score = 0.909 and sensitivity = 0.939, which are greater than U‐Net, CU‐Net, RCA‐IUnet and PMED‐Net.

GABAergic Inhibition Controls Receptive Field Size, Sensitivity, and Contrast Preference of Direction Selective Retinal Ganglion Cells Near the Threshold of Vision.

Information about motion is encoded by direction-selective retinal ganglion cells (DSGCs). These cells reliably transmit this information across a broad range of light levels, spanning moonlight to sunlight. Previous work indicates that adaptation to low light levels causes heterogeneous changes to the direction tuning of ON-OFF (oo)DSGCs and suggests that superior-preferring ON-OFF DSGCs (s-DSGCs) are biased toward detecting stimuli rather than precisely signaling direction. Using a large-scale multielectrode array, we measured the absolute sensitivity of ooDSGCs and found that s-DSGCs are 10-fold more sensitive to dim flashes of light than other ooDSGCs. We measured their receptive field (RF) sizes and found that s-DSGCs also have larger receptive fields than other ooDSGCs; however, the size difference does not fully explain the sensitivity difference. Using a conditional knock-out of gap junctions and pharmacological manipulations, we demonstrate that GABA-mediated inhibition contributes to the difference in absolute sensitivity and receptive field size at low light levels, while the connexin36-mediated gap junction coupling plays a minor role. We further show that under scotopic conditions, ooDSGCs exhibit only an ON response, but pharmacologically removing GABA-mediated inhibition unmasks an OFF response. These results reveal that GABAergic inhibition controls and differentially modulates the responses of ooDSGCs under scotopic conditions.

Industrial Product Surface Anomaly Detection with Realistic Synthetic Anomalies Based on Defect Map Prediction.

The occurrence of anomalies on the surface of industrial products can lead to issues such as decreased product quality, reduced production efficiency, and safety hazards. Early detection and resolution of these problems are crucial for ensuring the quality and efficiency of production. The key challenge in applying deep learning to surface defect detection of industrial products is the scarcity of defect samples, which will make supervised learning methods unsuitable for surface defect detection problems. Therefore, it is a reasonable solution to use anomaly detection methods to deal with surface defect detection. Among image-based anomaly detection, reconstruction-based methods are the most commonly used. However, reconstruction-based approaches lack the involvement of defect samples in the training process, posing the risk of a perfect reconstruction of defects by the reconstruction network. In this paper, we propose a reconstruction-based defect detection algorithm that addresses these challenges by utilizing more realistic synthetic anomalies for training. Our model focuses on creating authentic synthetic defects and introduces an auto-encoder image reconstruction network with deep feature consistency constraints, as well as a defect separation network with a large receptive field. We conducted experiments on the challenging MVTec anomaly detection dataset and our trained model achieved an AUROC score of 99.70% and an average precision (AP) score of 99.87%. Our method surpasses recently proposed defect detection algorithms, thereby enhancing the accuracy of surface defect detection in industrial products.

Focus Affinity Perception and Super-Resolution Embedding for Multifocus Image Fusion.

Despite the fact that there is a remarkable achievement on multifocus image fusion, most of the existing methods only generate a low-resolution image if the given source images suffer from low resolution. Obviously, a naive strategy is to independently conduct image fusion and image super-resolution. However, this two-step approach would inevitably introduce and enlarge artifacts in the final result if the result from the first step meets artifacts. To address this problem, in this article, we propose a novel method to simultaneously achieve image fusion and super-resolution in one framework, avoiding step-by-step processing of fusion and super-resolution. Since a small receptive field can discriminate the focusing characteristics of pixels in detailed regions, while a large receptive field is more robust to pixels in smooth regions, a subnetwork is first proposed to compute the affinity of features under different types of receptive fields, efficiently increasing the discriminability of focused pixels. Simultaneously, in order to prevent from distortion, a gradient embedding-based super-resolution subnetwork is also proposed, in which the features from the shallow layer, the deep layer, and the gradient map are jointly taken into account, allowing us to get an upsampled image with high resolution. Compared with the existing methods, which implemented fusion and super-resolution independently, our proposed method directly achieves these two tasks in a parallel way, avoiding artifacts caused by the inferior output of image fusion or super-resolution. Experiments conducted on the real-world dataset substantiate the superiority of our proposed method compared with state of the arts.

Learned Focused Plenoptic Image Compression With Microimage Preprocessing and Global Attention

Focused plenoptic cameras can record spatial and angular information of the light field (LF) simultaneously with higher spatial resolution relative to traditional plenoptic cameras, which facilitate various applications in computer vision. However, the existing plenoptic image compression methods present ineffectiveness to the captured images due to the complex micro-textures generated by the microlens relay imaging and long-distance correlations among the microimages. In this paper, a lossy end-to-end learning architecture is proposed to compress the focused plenoptic images efficiently. First, a data preprocessing scheme is designed according to the imaging principle to remove the sub-aperture image ineffective pixels in the recorded light field and align the microimages to the rectangular grid. Then, the global attention module with large receptive field is proposed to capture the global correlation among the feature maps using pixel-wise vector attention computed in the resampling process. Also, a new image dataset consisting of 1910 focused plenoptic images with content and depth diversity is built to benefit training and testing. Extensive experimental evaluations demonstrate the effectiveness of the proposed approach. It outperforms intra coding of HEVC and VVC by an average of 62.57% and 51.67% bitrate reduction on the 20 preprocessed focused plenoptic images, respectively. Also, it achieves 18.73% bitrate saving and generates perceptually pleasant reconstructions compared to the state-of-the-art end-to-end image compression methods, which benefits the applications of focused plenoptic cameras greatly. The dataset and code are publicly available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/VincentChandelier/GACN</uri> .

A Real-Time Contact-Free Atrial Fibrillation Detection System for Mobile Devices.

As the global population ages, the death and prevalence of atrial fibrillation (AF) continue to rise, posing significant concerns due to its strong association with stroke-related disabilities. Detecting AF early before a stroke occurs has become paramount. However, existing methods face challenges in achieving quick, easy, and affordable detection in complex environments characterized by motion interference and varying light conditions. To address these challenges, we propose a system that is employable for edge computing devices like smartphones, tablets, or laptops. Meanwhile, to ensure that the dataset reflects real-world scenarios, we collect 7,216 30-second segments from 452 subjects, categorized into Atrial Fibrillation (AF), Normal Sinus Rhythm (NSR), and Other Arrhythmias (Others), with a subject ratio of 105:116:231. Our lightweight non-contact facial rPPG atrial fibrillation detection system utilizes a Convolution Neural Network (CNN) with a large receptive field and a bidirectional spatial mapping augmented attention module (BiSME-ATT) coupled with a bidirectional feature pyramid network layer (BiFPN), optimized for deployment on mobile devices by reducing model parameters and floating-point operations per second (FLOPs). Our approach significantly improves AF detection accuracy, sensitivity, specificity, positive predictive value, and negative predictive value to 94.39%, 91.57%, 95.44%, 88.06%, and 96.93%, respectively, in AF vs. Non-AF scenarios. Furthermore, the results demonstrate notable enhancements in AF detection across various motion and light intensity levels.

LLIC: Large Receptive Field Transform Coding with Adaptive Weights for Learned Image Compression

LLIC: Large Receptive Field Transform Coding with Adaptive Weights for Learned Image Compression

Scale-Aware Super-Resolution Network With Dual Affinity Learning for Lesion Segmentation From Medical Images.

Convolutional neural networks (CNNs) have shown remarkable progress in medical image segmentation. However, the lesion segmentation remains a challenge to state-of-the-art CNN-based algorithms due to the variance in scales and shapes. On the one hand, tiny lesions are hard to delineate precisely from the medical images which are often of low resolutions. On the other hand, segmenting large-size lesions requires large receptive fields, which exacerbates the first challenge. In this article, we present a scale-aware super-resolution (SR) network to adaptively segment lesions of various sizes from low-resolution (LR) medical images. Our proposed network contains dual branches to simultaneously conduct lesion mask SR (LMSR) and lesion image SR (LISR). Meanwhile, we introduce scale-aware dilated convolution (SDC) blocks into the multitask decoders to adaptively adjust the receptive fields of the convolutional kernels according to the lesion sizes. To guide the segmentation branch to learn from richer high-resolution (HR) features, we propose a feature affinity (FA) module and a scale affinity (SA) module to enhance the multitask learning of the dual branches. On multiple challenging lesion segmentation datasets, our proposed network achieved consistent improvements compared with other state-of-the-art methods. Code will be available at: https://github.com/poiuohke/SASR_Net.

Chromatic processing and receptive-field structure in neurons of the anterior optic tract of the honeybee brain.

Color vision in honeybees is a well-documented perceptual phenomenon including multiple behavioral tests of trichromaticity and color opponency. Data on the combined color/space properties of high order visual neurons in the bee brain is however limited. Here we fill this gap by analyzing the activity of neurons in the anterior optic tract (AOT), a high order brain region suggested to be involved in chromatic processing. The spectral response properties of 72 units were measured using UV, blue and green light stimuli presented in 266 positions of the visual field. The majority of these units comprise combined chromatic-spatial processing properties. We found eight different neuron categories in terms of their spectral, spatial and temporal response properties. Color-opponent neurons, the most abundant neural category in the AOT, present large receptive fields and activity patterns that were typically opponent between UV and blue or green, particularly during the on-tonic response phase. Receptive field shapes and activity patterns of these color processing neurons are more similar between blue and green, than between UV and blue or green. We also identified intricate spatial antagonism and double spectral opponency in some receptive fields of color-opponent units. Stimulation protocols with different color combinations applied to 21 AOT units allowed us to uncover additional levels of spectral antagonism and hidden inhibitory inputs, even in some units that were initially classified as broad-band neurons based in their responses to single spectral lights. The results are discussed in the context of floral color discrimination and celestial spectral gradients.

DTFusion: Infrared and Visible Image Fusion Based on Dense Residual PConv-ConvNeXt and Texture-Contrast Compensation.

Infrared and visible image fusion aims to produce an informative fused image for the same scene by integrating the complementary information from two source images. Most deep-learning-based fusion networks utilize small kernel-size convolution to extract features from a local receptive field or design unlearnable fusion strategies to fuse features, which limits the feature representation capabilities and fusion performance of the network. Therefore, a novel end-to-end infrared and visible image fusion framework called DTFusion is proposed to address these problems. A residual PConv-ConvNeXt module (RPCM) and dense connections are introduced into the encoder network to efficiently extract features with larger receptive fields. In addition, a texture-contrast compensation module (TCCM) with gradient residuals and an attention mechanism is designed to compensate for the texture details and contrast of features. The fused features are reconstructed through four convolutional layers to generate a fused image with rich scene information. Experiments on public datasets show that DTFusion outperforms other state-of-the-art fusion methods in both subjective vision and objective metrics.

Seismic response prediction of a damped structure based on data-driven machine learning methods

Damping technology has been widely used because of its good vibration control effect. However, due to the strong nonlinearity of the added dampers, accurately predicting the seismic response of damped structures remains a challenge. This study investigates the application of interpretable machine learning (ML)-based and deep learning-based approaches to the prediction of the maximum inter-storey displacement of a damped structure. A comprehensive database consisting of 13,855 structural responses to ground motions was collected. Seven traditional interpretable ML algorithms including random forest and extreme gradient boosting (XGBoost), a convolutional neural network based on large receptive field, and seismic wave transformer (SWT) model based on a transformer network were developed. The predictions show that the error of the SWT based on unsupervised feature extraction is reduced by 50.90% compared with that of the optimal XGBoost in ensemble learning. Although the SWT has the highest global accuracy, XGBoost is found to have a smaller error when the structure is in a linear state with peak ground acceleration as partition index, so an aggregation model (AM)-based structural response prediction method was also proposed. The accuracy of the AM improved by 27.95% compared with that of the SWT. In contrast with other ML models, the proposed AM is more advantageous in terms of computational efficiency and accuracy.

WaveNet vocoder for prediction of time series with extreme events

Extreme events are typically defined as rare or unpredictable events that deviate significantly from typical behavior. Despite this, objective criteria for extreme events have yet to be established. Rareness may be characterized by certain scales or spatial and temporal boundaries, while intensity is an indication of an event’s potential to cause a significant change. One of the most prominent occurrences of extreme events in both neuroscience and medicine is in the case of epileptic seizures [1]. In speech synthesis, vocoder networks like WaveNet [2] generate audio. The model is a multi-layer convolutional neural network that functions as a causal filter and doesn’t predict the future. Due to this quality, the vocoder may have potential in time series prediction. Audio time series can be regarded as a dynamic system characterized by unpredictable switching regimes. For instance, transitioning from one letter to another can result in significant deviations in amplitude, similar to extreme events. This network receives r previous input counts known as a receptive field, and uses them to predict the next sample. The network is tree-like in structure, with exponentially increasing distances between subsequent layers of inputs. This is a necessary feature since the receptive field r is usually quite large, on the order of one or two thousand. Without this exponential increase in distance, the number of layers would depend linearly on r. Recurrent neural networks pose a challenge in optimizing the loss function when predicting time series sequences, as they tend to predict samples very similar to the previous one, causing the network to converge towards the mode. However, in a convolutional network, the output to the model will be longer due to the large receptive field. In the case of sound analysis, for instance, multiple oscillations occur within a given timeframe and the network does not elevate any specific sample. The study used artificial data generated from two coupled Hidmarsh–Rose neurons with chemical synaptic couplings. The observed variable was determined by the biological significance of the system, specifically the total membrane potential. The results exhibited extreme events across various coupling parameter values. Based on prior research [3], a numerical standard was selected for the events. The WaveNet vocoder model exhibits a 91% accuracy rate and 82% recall rate when forecasting extreme events of the same width as the prediction. It is noteworthy that recall is crucial in the forecast of extreme events since it identifies instances where the model predicted falsely that an extreme event would not occur.

Evolution of neuronal cell classes and types in the vertebrate retina

The basic plan of the retina is conserved across vertebrates, yet species differ profoundly in their visual needs1. Retinal cell types may have evolved to accommodate these varied needs, but this has not been systematically studied. Here we generated and integrated single-cell transcriptomic atlases of the retina from 17 species: humans, two non-human primates, four rodents, three ungulates, opossum, ferret, tree shrew, a bird, a reptile, a teleost fish and a lamprey. We found high molecular conservation of the six retinal cell classes (photoreceptors, horizontal cells, bipolar cells, amacrine cells, retinal ganglion cells (RGCs) and Müller glia), with transcriptomic variation across species related to evolutionary distance. Major subclasses were also conserved, whereas variation among cell types within classes or subclasses was more pronounced. However, an integrative analysis revealed that numerous cell types are shared across species, based on conserved gene expression programmes that are likely to trace back to an early ancestral vertebrate. The degree of variation among cell types increased from the outer retina (photoreceptors) to the inner retina (RGCs), suggesting that evolution acts preferentially to shape the retinal output. Finally, we identified rodent orthologues of midget RGCs, which comprise more than 80% of RGCs in the human retina, subserve high-acuity vision, and were previously believed to be restricted to primates2. By contrast, the mouse orthologues have large receptive fields and comprise around 2% of mouse RGCs. Projections of both primate and mouse orthologous types are overrepresented in the thalamus, which supplies the primary visual cortex. We suggest that midget RGCs are not primate innovations, but are descendants of evolutionarily ancient types that decreased in size and increased in number as primates evolved, thereby facilitating high visual acuity and increased cortical processing of visual information.