Group channel pruning and spatial attention distilling for object detection
نویسندگان
چکیده
Due to the over-parameterization of neural networks, many model compression methods based on pruning and quantization have emerged. They are remarkable in reducing size, parameter number, computational complexity model. However, most models compressed by such need support special hardware software, which increases deployment cost. Moreover, these mainly used classification tasks, rarely directly detection tasks. To address issues, for object network we introduce a three-stage method: dynamic sparse training, group channel pruning, spatial attention distilling. Firstly, select out unimportant channels maintain good balance between sparsity accuracy, put forward training method, introduces variable rate, rate will change with process network. Secondly, reduce effect propose novel method called pruning. In particular, divide into multiple groups according scales feature layer similarity module structure network, then use different thresholds prune each group. Finally, recover accuracy pruned an improved knowledge distillation Especially, extract information from maps specific as distillation. experiments, YOLOv4 PASCAL VOC dataset. Our reduces parameters 64.7% calculation 34.9%. When input image size is 416×416, compared original 256MB 87.1 accuracies, our achieves 86.6 accuracies 90MB size. demonstrate generality replace backbone Darknet53 Mobilenet also achieve satisfactory results.
منابع مشابه
On the Optimality of Spatial Attention for Object Detection
Studies on visual attention traditionally focus on its physiological and psychophysical nature [16, 18, 19], or its algorithmic applications [1, 9, 21]. We here develop a simple, formal mathematical model of the advantage of spatial attention for object detection, in which spatial attention is defined as processing a subset of the visual input, and detection is an abstraction with certain failu...
متن کاملOn-the-fly Network Pruning for Object Detection
Object detection with deep neural networks is often performed by passing a few thousand candidate bounding boxes through a deep neural network for each image. These bounding boxes are highly correlated since they originate from the same image. In this paper we investigate how to exploit feature occurrence at the image scale to prune the neural network which is subsequently applied to all boundi...
متن کاملislanding detection methods for microgrids
امروزه استفاده از منابع انرژی پراکنده کاربرد وسیعی یافته است . اگر چه این منابع بسیاری از مشکلات شبکه را حل می کنند اما زیاد شدن آنها مسائل فراوانی برای سیستم قدرت به همراه دارد . استفاده از میکروشبکه راه حلی است که علاوه بر استفاده از مزایای منابع انرژی پراکنده برخی از مشکلات ایجاد شده توسط آنها را نیز منتفی می کند . همچنین میکروشبکه ها کیفیت برق و قابلیت اطمینان تامین انرژی مشترکان را افزایش ...
15 صفحه اولMultiscale keypoint hierarchy for Focus-of-Attention and object detection
Hypercolumns in area V1 contain frequencyand orientation-selective simple and complex cells for line (bar) and edge coding, plus end-stopped cells for keypoint (vertex) detection. A single-scale (single-frequency) mathematical model of single and double end-stopped cells on the basis of Gabor filter responses was developed by Heitger et al. (1992 Vision Research 32 963-981). We developed an imp...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Applied Intelligence
سال: 2022
ISSN: ['0924-669X', '1573-7497']
DOI: https://doi.org/10.1007/s10489-022-03293-x