Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects
The majority of object detection algorithms based on convolutional neural network are focused on larger objects. In order to improve the accuracy and efficiency of small object detection, a novel lightweight object detection algorithm with attention enhancement is proposed in this paper. The network...
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2023-03-01
|
| Series: | Electronics |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2079-9292/12/7/1607 |
| _version_ | 1797608084448739328 |
|---|---|
| author | Nan Jia Zongkang Wei Bangyu Li |
| author_facet | Nan Jia Zongkang Wei Bangyu Li |
| author_sort | Nan Jia |
| collection | DOAJ |
| description | The majority of object detection algorithms based on convolutional neural network are focused on larger objects. In order to improve the accuracy and efficiency of small object detection, a novel lightweight object detection algorithm with attention enhancement is proposed in this paper. The network part of the proposed algorithm is based on a single-stage framework and takes MobileNetV3-Large as a backbone. The representation of shallower scale features in the scale fusion module is enhanced by introducing an additional injection path from the backbone and a detection head specially responsible for detecting small objects is added. Instead of pooling operators, dilated convolution with hierarchical aggregation is used to reduce the effect of background pixels on the accuracy of small object locations. To improve the efficacy of merging, the spatial and channel weights of scale features are modified adaptively. Last but not least, to improve the representation of small objects in the training datasets, the Consistent Mixed Cropping method is also proposed. The small labels of standard datasets are expanded with the self-collected samples for the training of the algorithm network. According to the test results and visualization on the 64-Bit Extended (X86-64) platform and embedded Advanced RISC Machine (ARM) platform, we find that the average accuracy (mAP) of the proposed algorithm is 4.6% higher than YOLOv4 algorithm, which achieves better small object detection performance than YOLOv4 algorithm, and the computational complexity is only 12% of YOLOv4 algorithm. |
| first_indexed | 2024-03-11T05:39:43Z |
| format | Article |
| id | doaj.art-4a99773df11e4e849cfeca63af8e388b |
| institution | Directory Open Access Journal |
| issn | 2079-9292 |
| language | English |
| last_indexed | 2024-03-11T05:39:43Z |
| publishDate | 2023-03-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Electronics |
| spelling | doaj.art-4a99773df11e4e849cfeca63af8e388b2023-11-17T16:33:03ZengMDPI AGElectronics2079-92922023-03-01127160710.3390/electronics12071607Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small ObjectsNan Jia0Zongkang Wei1Bangyu Li2Beijing Institute of Aerospace Control Device, Beijing 100854, ChinaBeijing Institute of Aerospace Control Device, Beijing 100854, ChinaInstitute of Automation Chinese Aacademy of Sciences, Beijing 100098, ChinaThe majority of object detection algorithms based on convolutional neural network are focused on larger objects. In order to improve the accuracy and efficiency of small object detection, a novel lightweight object detection algorithm with attention enhancement is proposed in this paper. The network part of the proposed algorithm is based on a single-stage framework and takes MobileNetV3-Large as a backbone. The representation of shallower scale features in the scale fusion module is enhanced by introducing an additional injection path from the backbone and a detection head specially responsible for detecting small objects is added. Instead of pooling operators, dilated convolution with hierarchical aggregation is used to reduce the effect of background pixels on the accuracy of small object locations. To improve the efficacy of merging, the spatial and channel weights of scale features are modified adaptively. Last but not least, to improve the representation of small objects in the training datasets, the Consistent Mixed Cropping method is also proposed. The small labels of standard datasets are expanded with the self-collected samples for the training of the algorithm network. According to the test results and visualization on the 64-Bit Extended (X86-64) platform and embedded Advanced RISC Machine (ARM) platform, we find that the average accuracy (mAP) of the proposed algorithm is 4.6% higher than YOLOv4 algorithm, which achieves better small object detection performance than YOLOv4 algorithm, and the computational complexity is only 12% of YOLOv4 algorithm.https://www.mdpi.com/2079-9292/12/7/1607small object detectionattention mechanismlightweightembed inference test |
| spellingShingle | Nan Jia Zongkang Wei Bangyu Li Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects Electronics small object detection attention mechanism lightweight embed inference test |
| title | Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects |
| title_full | Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects |
| title_fullStr | Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects |
| title_full_unstemmed | Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects |
| title_short | Attention-Enhanced Lightweight One-Stage Detection Algorithm for Small Objects |
| title_sort | attention enhanced lightweight one stage detection algorithm for small objects |
| topic | small object detection attention mechanism lightweight embed inference test |
| url | https://www.mdpi.com/2079-9292/12/7/1607 |
| work_keys_str_mv | AT nanjia attentionenhancedlightweightonestagedetectionalgorithmforsmallobjects AT zongkangwei attentionenhancedlightweightonestagedetectionalgorithmforsmallobjects AT bangyuli attentionenhancedlightweightonestagedetectionalgorithmforsmallobjects |