Abstract:The convolutional neural networks (CNNs) deployed on devices for visual image processing faces the thorny problems on high system real-time requirements and resource consumption. A high-performance Low-res MobileNet model is constructed to effectively alleviate the high computing resources and storage costs in the real-time image processing. The main works are summarized as: (1) To actively match the input of low-resolution feature map, the MobileNetV2 is further optimized by clipping to simplify the network structure and improve the efficiency of image recognition. (2) To improve the classification accuracy, the Inception structure is used to fill the Dwise layer in depthwise separable convolution to extract more abundant low-resolution features; the activation function during the process of increasing the dimension is replaced to avoid the loss of useful information; Inter-layer connection structure is adopted to strengthen the fusion of feature information between layers. (3) To reduce the network scale, the gradually decreasing expansion factors are used to remove the redundant structure of the model. Subsequently, the Low-res MobileNet is validated and evaluated through data sets of different scales. The experimental results show that this model has smaller scale, less computation and higher classification accuracy compared with other CNN models. The model has 0.36 M parameters and 25.46 M floating point of operations (FLOPs), which is easy to deploy to resource-constrained mobile and embedded devices. The model runs at 35 batches per second, and it achieves an accuracy rate of 89.38%, 71.60%, and 87.08% on CIFAR-10, CIFAR-100, and CINIC-10 datasets, respectively, which is basically suitable for real-time image classification task applied in low-resolution application scenarios.

DenseNet Models for Tiny ImageNet Classification

Image classification based on tensor network DenseNet model

Connection Reduction of DenseNet for Image Recognition

A Tensor-based Convolutional Neural Network for Small Dataset Classification

SparseNet: A Sparse DenseNet for Image Classification

Low-res MobileNet: An efficient lightweight network for low-resolution image classification in resource-constrained scenarios

Efficient densely connected convolutional neural networks

SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size

SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size

T3DNet: Compressing Point Cloud Models for Lightweight 3D Recognition

Model Rubik's Cube: Twisting Resolution, Depth and Width for TinyNets

AttendNets: Tiny Deep Image Recognition Neural Networks for the Edge via Visual Attention Condensers

A New Image Classification Approach via Improved MobileNet Models with Local Receptive Field Expansion in Shallow Layers

Multi-Scale Dense Networks for Resource Efficient Image Classification

Convolutional Networks with Dense Connectivity

Tiny Object Detection Model Based on Competitive Multi-Layer Neural Network (TOD-CMLNN)

Efficient Classification of Very Large Images with Tiny Objects

Dense Transformer Networks

When Residual Learning Meets Dense Aggregation: Rethinking the Aggregation of Deep Neural Networks

Classification of Astronomical Bodies by Efficient Layer Fine-Tuning of Deep Neural Networks

Enhancing Image Classification Accuracy Based on AlexNet