The correct identification of timber species is a complicated task for the wood industry and government institutions regulating the different laws that ensure legal and transparent commerce. Currently, experts perform this process using the organoleptic characteristics of the wood. However, the methodology used is time-consuming and limited to environmental conditions. Moreover, it has a scalability issue since acquiring this specific knowledge and experience has a slow learning curve. On the other hand, deep learning models have evolved as possible solutions for process automation. Therefore, this paper explores convolutional neural network models suited to run on edge devices. The present study created a database with 25k images of 25 timber species from the Peruvian Amazon. We trained-validated multiple lightweight models (less than 5M). The experiments were made using a repeated stratified k-fold cross-validation approach to estimate the performance of the classifiers. The experiments show that the best model has an F1 score metric of 99.90% and 58ms latency using 561k parameters. Furthermore, the created model showed an excellent ability to identify species, opening up space for future integration with mobile applications, which helps minimize the time spent and the identification errors on timber identification carried out by experts on control points.
This study used correctly identified wood core pieces and timber images from principal commercial species in Peru to construct the dataset used to train and validate the different convolutional neural network models. The timber images were collected from Serfor trucks control transportation checkpoints located in La Merced, Chanchamayo Province in the Jun ́ın Region in Peru. A cutting knife, cell phone, and a portable microscope were used to capture multiple images from different timber species. The portable microscope acquired images with a magnification of 10x. No special treatment (e.g., pores filling or sanding) was applied to wood samples.
The images to train and validate and its metadata can be dowloaded to replicate the results.
Neural networks (NN) applied to computer vision are mainly based on the concept of receptive fields, which enable the network to explore spatial correla- tions in the image. As a result, neural networks are able to transform the crude information in an input image into a more meaningful representation of its content. However, nowadays, NN are growing larger and larger, which makes them difficult to be used on edge devices since the existing hardware limita- tions. Thus the use of a few parameter models is more suitable to work
under those devices conditions, for this NN has different particular components which make them unique:
-
Residual bottleneck: Wide-narrow-wide structure with the number of chan- nels. The input has a high number of channels, which are compressed with a 1x1 convolution. The number of channels is then increased again with a 1x1 convolution so input and output can be added
-
Inverted residual bottleneck: A narrow-wide-narrow approach, hence the inversion. First, input is widened with a 1x1 convolution. Then a 3x3 depthwise convolution is used (which significantly reduces the number of parameters). Finally, a 1x1 convolution is used to reduce the number of channels so input and output can be added.
-
Depthwise convolutions: Type convolution where a single convolutional filter is applied for each input channel.
in_channels = 3
depth_conv = Conv2d(in_channels=in_channels, out_channels=in_channels, kernel_size=3, groups=in_channels)- Pointwise convolutions: Type of convolution that uses a 1x1 kernel. This kernel iterates through every single point and has a depth of however many channels the input image has. It can be used in conjunction with depthwise convolutions to produce an efficient class of convolutions known as depthwise-separable convolutions.
in_channels, out_channels = 3, 9
point_conv = Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1)In this work, we focus on the following models: MobileNetV2, EfficientNet-B0, Residual NN with patches
Look a the folder train_executions for training examples with fold partirion and full images.
The results of the experiments are summarised in the table below. The F1-Scores were scaled by 100 to contrast the values.
| Model | fold-1 | fold-2 | fold-3 | fold-4 | test |
|---|---|---|---|---|---|
| Patches-ResNet | 99.24 | 98.97 | 99.32 | 99.41 | 99.90 |
| MobileNetV2 | 99.29 | 99.49 | 99.43 | 99.21 | 99.50 |
| EfficientNet-B0 | 99.24 | 99.01 | 99.34 | 99.38 | 99.60 |
| EfficientNet-B0-NS | 99.21 | 99.08 | 99.05 | 98.97 | 92.60 |
| RestNet | 98.91 | 98.81 | 98.94 | 98.63 | 99.00 |